Abstract

Breast cancer is the most frequently diagnosed cancer in women all around the world, with approximately 2.3 million cases arising globally each year. Breast cancer remains a leading cause of cancer-related morbidity and mortality worldwide. Integrating artificial intelligence (AI) into breast cancer detection and treatment brings about potential improvements in risk assessment, treatment personalisation, and early detection. This paper examines and analyses various case studies to illustrate the application of AI in these crucial domains. We aim to determine the advantages, difficulties, and future research directions for integrating AI into breast cancer management. The case studies demonstrate how AI algorithms increase mammography accuracy by decreasing false positives by 5.7% and false negatives by 9.4%. AI proved to improve diagnostic efficiency by reducing interpretation time by roughly 20-30% and improving patient outcomes by 10%-15% by more precisely customising therapies to individual profiles. 

These real-world applications theoretically validate AI’s capacity to improve disease models and extract knowledge from large data sets. Through this evaluation, we hope to provide a comprehensive overview of how AI can be integrated into clinical practice, improving patients’ lives and furthering the field’s research. To fully realise AI’s potential benefits, future research should address ethical issues, incorporate AI into clinical workflows, and assess the long-term impact on patient outcomes.

Introduction

Breast cancer is a disease where cells in the breast tissue proliferate uncontrollably, forming a mass or tumour (World Health Organisation, 2024). Inside the breast’s milk ducts and/or milk-producing lobules are where breast cancer cells first proliferate. The earliest form (in situ) is non-lethal and can be detected early on. Invasive cancers can metastasise to nearby lymph nodes or other organs. Metastasis can be lethal and potentially fatal (World Health Organisation, 2024). Late-stage diagnosis and treatment variability continue to be major challenges despite advancements in diagnostic technologies and treatment strategies. With its creative approaches to improving breast cancer treatment, artificial intelligence (AI) has become a groundbreaking instrument in the healthcare industry. AI is a general term for a number of technologies, such as machine learning (ML), deep learning (DL), and natural language processing (NLP), which are intended to analyse large, complicated data sets and yield useful insights (Coursera.org, 2024). 

Machine Learning (ML)

As the name suggests, machine learning (ML) is the process of computers learning on their own without explicit programming or human assistance. The first step in the machine learning process is to feed the machines high-quality data. The machines are then trained by creating different machine-learning models with various algorithms and data. The algorithms we use are determined by the kind of task we are trying to automate and the type of data we have (Sarker, 2021).

Deep Learning (DL)

Deep learning enables computational models made up of multiple processing layers to learn representations of data at various levels of abstraction. DL has significantly advanced the abilities of state-of-the-art technologies in speech recognition, visual object recognition, object detection, and a variety of other fields, including drug discovery and genomics. By using the backpropagation algorithm to suggest changes to a machine’s internal parameters that are used to compute the representation in each layer based on the representation in the previous layer, deep learning uncovers complex structure within massive data sets (LeCun et al., 2015).

Natural Language Processing (NLP)

The goal of the computer science and artificial intelligence subfield of natural language processing (NLP) is to enable computers to comprehend human language. NLP uses several models based on statistics, machine learning, deep learning, and computational linguistics (the study of language). Thanks to these advancements, computers are now able to process and analyse text or voice data and fully understand what is being said or written, including the intentions and feelings of the speaker (Deeplearning.AI, 2024).

The objective of this review is to investigate the integration of AI in breast cancer treatment, with a focus on its applications in early detection, personalised treatment planning, and predictive modelling. By analysing specific case studies, this study aims to evaluate the effectiveness and challenges of AI technologies in enhancing clinical outcomes for breast cancer patients. The findings hold significant potential for advancing the development of more accurate diagnostic tools and personalised therapies, ultimately improving patient care and survival rates in breast cancer treatment.

Integration of Certain AI Tools in the Treatment of Breast Cancer: Recent Developments

The exploration of these additional AI tools serves to complement the insights gained from the case studies by showcasing a wider array of technological advancements. This broader examination highlights the versatility and adaptability of AI in breast cancer treatment, offering a more complete understanding of how these innovations are shaping the future of patient care and driving progress in the field. These developments provide further proof of how AI is being implemented in the medical field as time progresses.

AI in Radiomics

The integration of AI in radiological imaging is rapidly expanding. Radiomics is the process of extracting features from images such as shape, size, and texture. These characteristics can be used in an algorithm to help diagnose various conditions. Radiomics can be used for a wide range of applications, including screening, tumour burden prediction, and therapy guidance. It helps to classify nodules, describe tumours, predict metastatic potential, and respond to treatments (Patel et al., 2011). 72% of healthcare providers think artificial intelligence can automate at least 20% of their administrative workload (Medtech Intelligence, 2024). This means prolonged time for consultations, improved results for patients, and reduced burnout among professionals. 

Virtual Assistants

AI virtual assistants can collect and analyse patient data from various sources, including appointment scheduling patterns, medication adherence, and symptom logs. Ezra (a company that offers full-body MRI screening services to help detect potential health issues early) is working on building an artificial intelligence system that will reduce analysis time, while increasing radiologist accuracy. In a statement released by Ezra in Medium (2024), they state: “Our mission at Ezra is to detect cancer early for everyone worldwide. We plan to do so by offering fast, accurate, pain-free MRI-based cancer screening; provide access to expert radiologists, assisted by the Ezra AI (once we obtain FDA clearance).”

MammoScreen

Therapixel’s MammoScreen is a prime example of a tool that could significantly enhance the efficiency of breast cancer detection in the future. This innovative tool analyses various regions of the breast, highlighting areas that may require further examination. By doing so, MammoScreen not only simplifies the process of identifying potential lesions, but also reduces the time it typically takes a radiologist to detect cancer.

MammoRisk

MammoRisk is another AI-based tool designed to assess a person’s risk of developing invasive breast cancer. This tool evaluates risk by considering five key factors: age, family history, biopsy history, polygenic scores, and breast density (Predilife, 2024). Based on this evaluation, MammoRisk generates a detailed report for the doctor, allowing them to tailor treatment and preventive measures according to the patient’s individual risk profile.

Mammography Intelligent Assessment (MIA)

When reviewing imaging, doctors can sometimes miss potential signs of cancer. The tool Mammography Intelligent Assessment (MIA) helps to make the radiologist’s job easier by detecting signs of cancer that might be overlooked by the human eye. According to a study, MIA enabled doctors to identify 12% more cancers than usual, while also reducing their workload by 30% (Microsoft UK Stories, 2024).

The integration of AI into radiological imaging is revolutionising the way we detect, diagnose, and manage various health conditions. Tools like AI in Radiomics, MammoScreen, MammoRisk, and MIA are not only enhancing diagnostic accuracy, but also reducing the workload on healthcare professionals, leading to better patient outcomes and more efficient care. As AI continues to evolve, its role in healthcare will undoubtedly expand.

In the next section, we will delve into specific case studies that highlight the real-world impact of these AI-driven innovations, examining how they have been applied in clinical settings to improve patient care and outcomes. We conducted a thorough analysis of three case studies that focus on risk assessments, treatment options, and preventative care in order to gain an understanding of the effectiveness, challenges, and future potential of AI technologies in regards to breast cancer care. 

Case Study 1: AI Use in the Early Detection of Cancer

This case study describes an extensive strategy to create and assess an advanced AI system for breast cancer detection based on ultrasound images. The AI system was designed and tested using the NYU Breast Ultrasound Dataset41, which contained 5,442,907 images from 288,767 breast ultrasound exams performed on 143,203 patients between 2012 and 2019 at NYU Langone Health in New York, USA (Shen et al., 2021).

The data set included 28,914 exams with a pathology report, and the biopsy or surgery revealed benign or malignant results for 26,843 and 5,593 patients respectively. The patients in the data set were split into three groups at random: a validation set (10%), which was used for hyperparameter tuning; an internal test set (30%), which was used to test the accuracy of the AI system in recognising cancerous tumours; and a training set (60%), which was used for model training (Shen et al., 2021).

Using the ultrasound images and related pathology reports as training data, the AI system was trained to recognise cancerous tumours. Through an in-depth reader study, the system’s performance was compared against ten board-certified radiologists. To determine whether this strategy may further improve diagnostic accuracy, the researchers also created a hybrid model that incorporated both the predicted results made by the AI system and the opinions of the radiologists.

Binary classification tasks use the performance metric known as Area Under the Receiver Operating Characteristic curve, or AUROC. The true positive rate (sensitivity) is plotted against the false positive rate (1-specificity) at different threshold settings to determine how well a model can distinguish between the two classes. The AUROC value ranges from 0 to 1; a value closer to 1 indicates better model performance, while 0.5 represents random guessing. A higher AUROC value indicates that the model is effective at distinguishing between the positive and negative classes (Janssens & Martens, 2020). 

The study compared the performance of an AI system with breast radiologists using an internal test set and an external BUSI data set (Al-Dhabyani et al., 2020). In the internal data set of 44,755 US exams (25,003 patients, 79,156 breasts), the AI mechanism obtained an AUROC of 0.976 for detecting breasts with cancerous lesions. Patients were then categorised by age and mammographic breast density, and AI model performance was assessed across these sub-populations. The system demonstrated high diagnostic accuracy across all age groups (AUROC: 0.969-0.981) and mammographic breast densities (AUROC: 0.964-0.979). Furthermore, it was investigated how the size of the training data set affected the AI system’s performance. It was observed that more training data resulted in a higher AUROC, indicating that this system can improve further with time as more data is included in the training sets. In addition, the AI system was tested on an external test set (the BUSI data set). Despite not being trained on any images from the external test set, the AI system maintained a high level of diagnostic accuracy (AUROC: 0.927, 95% CI: 0.907, 0.959) (Shen et al., 2021).

The internal test set consisted of 663 exams, with 73 breasts having pathology-proven cancer, 535 having benign findings, and 416 breasts not biopsied but evaluated as likely benign. Ten board-certified breast radiologists rated each breast according to the Breast Imaging Reporting and Data System (BI-RADS). The AI system achieved a higher AUROC and AUPRC (Area Under the Precision-Recall Curve), which is another measure of classification accuracy, compared to the average radiologist. The AI system also achieved improved specificity and sensitivity compared to radiologists, with the AI system recommending tissue biopsies on 19.8% of breasts and 32.5% of these biopsies being for breasts ultimately found to have cancer. It outperformed radiologists in terms of sensitivity (94.5%) and specificity (85.6%), while also reducing unnecessary biopsies by 4.5%, and increasing positive predictive value (PPV) by 5.4%. Overall, the AI system performed better across multiple metrics, indicating that it could significantly improve breast cancer detection, reduce false positives, and lead to better patient outcomes compared to when these processes rely solely on radiologists (Shen et al., 2021).

Although promising, the AI system for breast ultrasound interpretation has several limitations, including data set precision, which could lead to bias, potential overfitting, medical integration challenges, issues handling rare cases, and a lack of comprehension and transparency. Concerns have also been raised about the long-term effects on patient results and medical expenses. Regulatory and ethical issues regarding patient privacy and data security may arise as well (Lotter et al., 2020).

However, we believe that these problems can be addressed. For example, worldwide collaborative databases can be used to diversify training data, and regularisation and cross-institutional validation strategies can be used to reduce overfitting. Clinical integration can be made easier through user-centric design. Continuous learning mechanisms and hybrid AI-human systems are essential for edge cases. Explainable AI models and interactive decision support can improve transparency, whereas longitudinal studies and cost-benefit analyses can contribute to effective long-term implementation. Regulatory frameworks and ethics review boards should oversee the responsible use of AI systems.

Therefore, through this case study we can determine that proper use of artificial intelligence in early breast cancer detection can transform diagnostic practices by improving accuracy, reducing diagnostic delays, and ultimately improving the lives of patients through timely and precise intervention.

Case Study 2: Use of AI in Personalised Treatment Planning

The Dana-Farber Cancer Institute case study displays AI’s transformative potential in personalised medicine for breast cancer treatment. The organisation made use of IBM Watson for Oncology (WFO), a cutting-edge AI tool created to help with the creation of customised treatment regimens based on a patient’s distinct genetic profile and medical information. The main objective of this AI tool is to enhance the accuracy of treatment recommendations through the integration of various complex data sources, such as genetic mutations, medical history of the patient, and specific tumour characteristics. Because Watson for Oncology can process and analyse large data sets, it can provide highly personalised treatment options for each patient. It would typically take doctors weeks to develop a personalised treatment plan, but Watson can analyse patients’ genetic profiles and provide options in a matter of minutes. The program is the most recent effort in IBM’s larger Watson Health plan aimed at enhancing healthcare. The AI tool was trained by The Memorial Sloan Kettering Cancer Center (MSKCC) and was first introduced in 14 different countries (Hamilton et al., 2019). Watson employs machine learning to determine how to weigh clinical factors in a patient’s case in order to identify treatment options based on the data that has been processed. Training refers to the process of developing and tuning the algorithms that Watson uses to provide decision support (Malin, 2013). ) 

In order to create a customised treatment plan, doctors use the supercomputer’s sophisticated cognitive powers to interpret DNA insights, comprehend a patient’s genetic profile, and compile data from pertinent medical literature. Watson’s rationale and insights will improve as it is used more, allowing it to provide physicians with the most recent combined wisdom from national oncology leaders. In a recent large-scale study, researchers at the Manipal Comprehensive Cancer Centre in Bengaluru, India, found that, on average, 90% of Watson recommendations or recommendations for consideration agreed with the tumour board’s recommended course of treatment for 638 patients with breast cancer. Upon comparing the duration required to gather and evaluate data “manually” with Watson’s ability to produce recommendations, the researchers found that Watson produced recommendations significantly quicker. They discovered that the manual method required an average of twenty minutes; this was lowered to twelve minutes as doctors became more experienced with the cases. Meanwhile, Watson analysed the data in just 40 seconds and made a treatment recommendation (Cavallo, 2017).

This case study further explores the effectiveness of the IBM Watson for Oncology AI tool by comparing cancer treatment options proposed by the IBM Watson for oncology (Artificial Intelligence), and the usual clinical practice for cancer patients in China.Data from 362 cancer patients were entered into WFO between April and October 2017. The recommendations for treatment were provided in three distinct categories: recommended, for consideration, and not recommended. The treatment plans of both the WFO and the Chinese physicians were compared. Concordance was achieved when both the WFO and oncologists’ plan was either in the recommended or for consideration category (Zhou et al., 2020).

The participants consisted of: 113 (31.2%) patients with lung cancer, 120 (33.1%) with breast cancer, 42 (11.6%) with gastric cancer, 25 (6.9%) with colon cancer, 24 (6.6%) with cancer, 14 (3.9%) with cervical cancer, and 24 (6.6%) with ovarian cancer (Zhou et al., 2020)

Out of 113 lung cancer patients, 22% had SCLC (small cell lung cancer), and 78% had NSCLC (non-small cell lung cancer). The patients were also concurrently distributed into groups based on histology and tumour stage. Out of 113 patients, for 93 the treatment path was concordant between WFO and the physicians (81.3%). Concordance for both types of lung cancer: SCLC 92%, NSCLC 79.99%. Concordance for specific cancer stages: 2nd stage lung cancer 87.5%, 3rd stage lung cancer 75.8%, 4th stage lung cancer 84.6%.(Zhou et al., 2020)

The treatment plans were concordant in 65% of cases of second stage breast cancer, and in 64.1% of cases for third stage breast cancer. There was only one patient with stage four breast cancer, for whom the physician chose the treatment plan (Zhou et al., 2020).

The accuracy of cancer treatments may be impacted by problems with IBM Watson for Oncology at Dana-Farber, such as integrating incoherent patient data and comprehending unorganised data. We believe that implementing robust data standardisation, improving Watson’s natural language processing to better handle complex cases, and developing a continuous learning system to keep up with the latest research are all potential solutions. Increasing the transparency of AI decision-making and offering trust-building education are two ways to combat physician scepticism. To reduce costs, a phased implementation strategy, cloud-based solutions, and seeking external funding should be considered.

We believe that AI can be very helpful in modern medicine in treatment planning and other fields of medicine. However, artificial intelligence can only be implemented into healthcare if we develop it by including more varied data, such as data from less medically developed and underrepresented countries. For now, we can see massive potential in AI, and it’s just a matter of time before it comes to the healthcare market.

Case Study 3: AI in Predictive Risk Assessment

The Polygenic Risk Score (PRS) model identifies the risk a patient has in getting breast cancer by analysing the individual’s genetic data, using AI. This model has been expertly developed by the Karolinska Institute, which is globally renowned for its significant contributions in medical research. The model creates a significant advancement in breast cancer research. The PRS model, by classifying patients into numerous categories based on their risk, may be able to revolutionise breast cancer treatment (Edgren et al., 2007), by aiding in the early identification of breast cancer. This can result in the saving of lives of countless people at risk of breast cancer.

Annually, around 2.4 million individuals are diagnosed with breast cancer, making this disease a significant health challenge. By being able to detect breast cancer in its earlier stages, doctors are able to increase survival rates and treat patients more effectively. Alternative risk assessment tools, like mammograms, work well but are often inadequate in their accuracy in identifying which patients are more likely to develop the disease. Genetic testing through AI has created several opportunities for risk assessment, and the PRS model stands to be one of the most promising ones (Torkmani et al., 2018).

The PRS model quantifies the patient’s genetic predisposition to breast cancer by identifying genetic variants commonly found in breast cancer patients. Unlike monogenic diseases (which are caused by a single gene mutation), one’s risk of developing breast cancer is affected by multiple genetic mutations and factors. With the PRS model, these variants are grouped into clusters and then into a single score, which creates a complete risk profile of the individual (Mavaddat  et al., 2019).

The PRS model utilises the data found in the genome-wide association studies (GWAS), which are used to identify variants (also known as single nucleotide polymorphisms (SNPs) which are associated with the disease. For breast cancer, GWAS have identified several SNPs that are linked to the development of the disease (Manolio et al., 2009). The PRS model first collects genetic data from thousands of patients, with and without a diagnosis of the disease (Edgren et al., 2007). Following this, the data is thoroughly analysed to recognise SNPs that are commonly seen in patients with known breast cancer. Then, the model identifies how strongly each SNP contributes to the risk of breast cancer, so that doctors can pay more attention to them. Through statistical models, the PRS model identifies the frequency of the SNP and its associated risk. Finally, the model identifies the number of breast cancer-related SNPs in an individual to then classify their genetic predisposition to breast cancer (Torkamani et al., 2018).

The PRS model is able to analyse large amounts of SNPs, however small their risk may be to the patient. Traditional testing tools only identify the higher-risk variants, such as those with the BRCA1 and BRCA2 genes. However, the SNPs the traditional tools recognise are relatively rare in most patients, making the PRS model preferable. By considering the effect that all the variants can have, the model can then predict for a larger population (Torkamani et al., 2018).

Overall, the goal of this model is to apply it into clinical settings to enhance the prevention of breast cancer in early stages. After this tool has been validated, it can be utilised by doctors globally, allowing them to stratify patients into different risk clusters based on their genetic data. This stratification then leads to tailored preventive measures of a patient and formation of a personal profile for healthcare workers to utilise when scheduling screenings.

The most crucial impact of the model is the potential to enhance the decision-making of healthcare workers, by providing a more holistic understanding of the risk an individual has. This can result in more personalised treatment and screening approaches. Women who have been identified as high-risk by PRS will have the opportunities for more forward screening and prevention, and are likely to be advised to undergo a mammogram. These patients can benefit immensely from early action strategies and preventive measures, such as prophylactic mastectomy, or the use of tamoxifen or raloxifene, which are commonly used to reduce the risk of breast cancer in women (Torkamani et al., 2018). They could also have more treatment opportunities, such as selective oestrogen receptor modulators (SERMs) or prophylactic surgery. In contrast, women that have not been identified as high-risk by PRS will be advised to continue following standard screening schedules, to clear any potential anxiety and costs that come with unnecessary tests (Manolio et al., 2009). The model therefore allows the elimination of over-screening, unnecessary spending on biopsies, and especially false positives (Pharaoh et al., 2002). By altering prevention strategies to adhere to an individual’s genetic predispositions, healthcare providers can focus on the treatment opportunities that fit the patient’s disease and budget (Pharaoh et al., 2002). 

Aside from tailored care, the model can also have broader public health implications. The model allows targeted interventions directed at the public to reduce the mortality rates of breast cancer. For example, these campaigns could educate women on the effects of breast cancer and suggest they use the model to identify their risk for the disease. This would allow more women to understand their genetic predispositions so that they can take early action. 

The primary limitation of the PRS model is its reliance on genetic data which has primarily been collected from European populations. GWAS predominantly focuses on European ancestry, meaning the model may not be helpful or accurate when treating a patient with differing ancestry. This flaw keeps millions of women from being able to identify their genetic mutations, leading to many undiagnosed and untreated patients (Torkamani et al., 2018). To address this specific issue, the GWAS must expand the data it uses to include populations from different parts of the world. Through global efforts to collect the essential genetic data from minority populations, the PRS model can be much more accurate in its findings. The development of specific models to different populations is another solution. For example, different PRS models could be modified to represent different backgrounds to take the different genetic variants in each culture into consideration.

An individual’s risk does not rely only on genetic mutations, but also  on one’s environmental and lifestyle factors, such as exercise, diet, illicit substance/alcohol abuse, and exposure to different cancer-causing substances (carcinogens). The PRS model is incapable of taking these things into consideration since it only focuses on the genetic risk of a patient. Therefore, using PRS can oversimplify one’s risk of breast cancer, by failing to detect the contributions that non-genetic factors have on the development of breast cancer. To avoid this, the environmental data can be implemented based on the area the machine is being used in. The machine can also take data from other parts of the patient aside from its genes to see if they lead to the disease (Manolio et al., 2009). 

The use of genetic data in risk prediction creates serious ethical and privacy concerns. Genetic discrimination is highly likely where individuals may not get proper treatment solely based on their genetic information. For example, individuals with greater risk may be turned down by insurance companies or employment opportunities. Furthermore, the security of a patient’s genetic information is crucial, and highlighting it may have detrimental effects on the patient’s privacy and autonomy. To prevent this, it is absolutely vital to implement complex data security protocols to ensure the machine functions in line with ethical guidelines. This must include consent from the patient before their genetic data is collected to ensure their knowledge on the potential risks of participating in genetic studies. The Genetic Information Nondiscrimination Act (GINA) in the United States is a primary example that other countries should follow when establishing the ethical collection and use of genetic information (Genetic Information Nondiscrimination Act, 2018). Finally, public campaigning is an effective way to raise awareness on the negative effects genetic testing has on the ethics in healthcare. 

Overall, the Polygenic Risk Score (PRS) model provides an example of the significant enhancements AI can have on predictive medicine. Utilising genetic data creates an individualised approach to breast cancer prevention and potentially saves thousands of lives by detecting the disease in its earliest stages.

Conclusion

In conclusion, the case studies we have examined demonstrate how significantly artificial intelligence (AI) can improve early detection, plan individualised treatments, and predict patient risk, when it comes to breast cancer care. The NYU Langone Health study showed that AI is capable of reducing false positives and outperforming radiologists in sensitivity, which could result in earlier diagnoses and therefore more promising patient results. As an example of how AI can closely align with expert decisions, the Dana-Farber Cancer Institute used the technology for personalised treatment planning, enabling more accurate and quicker recommendations of therapies. The predictive risk assessment research conducted by the Karolinska Institute further demonstrated AI’s ability to precisely classify risk, opening the door for focused preventive measures. Additional AI tools such as Virtual Assistants, MammoScreen, and MammoRisk can also contribute to the improvement of breast cancer care. For example, Mammogram Intelligent Assessment (MIA) has improved ultrasound imaging.

For AI to be widely and smoothly implemented it will be crucial to solve existing issues like data diversity, model transparency, and integration into clinical workflows as the technology develops. While AI has enormous potential, its successful implementation will require concerted efforts to overcome existing barriers. In our opinion, experts should place a high priority on diversifying training data sets to guarantee that AI models can be applied to a variety of demographics, which will lessen bias and enhance accuracy for under-represented groups. Furthermore, promoting transparency in AI models is critical; this can be accomplished by creating interpretable AI systems that provide clear, actionable insights that clinicians can trust and comprehend. Another significant issue is integration into clinical workflows; experts should concentrate on creating AI systems that enhance current procedures rather than cause them to change. To create user-friendly interfaces that blend in with daily routines, radiologists, oncologists, and AI developers may collaborate across disciplines. With continued innovation and collaboration, AI holds the promise of saving countless lives and elevating the quality of healthcare worldwide.

Ahn, J.S., Shin, S., Yang, S.A., Park, E.K., Kim, K.H., Cho, S.I., Ock, C.Y., & Kim, S. (2023) Artificial intelligence in breast cancer diagnosis and personalised medicine. Journal of Breast Cancer, 26(5), p.405-435. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC10625863/; https://pubmed.ncbi.nlm.nih.gov/38554069/ 

Al-Dhabyani, W., Gomaa, M., Khaled, H., & Fahmy, A. (2020). Dataset of breast ultrasound images. Data in Brief, 28, 104863. https://doi.org/10.1016/j.dib.2019.104863

Bender, E. (2020). Bringing AI towards personalised cancer care. Dana-Farber Cancer Institute. https://www.dana-farber.org/newsroom/publications/paths-of-progress-2020/bringing-ai-toward-personalized-cancer-care

Cavallo, J. (2017).  How Watson for Oncology is advancing personalised patient care. The ASCO Post. https://ascopost.com/issues/june-25-2017/how-watson-for-oncology-is-advancing-personalized-patient-care/ 

Coursera (2024) What is artificial intelligence? Definition, uses, and types. https://www.coursera.org/articles/what-is-artificial-intelligence?msockid=28f7073127b96e85226e0893260b6fb7

DeepLearning.AI (2023). Natural Language Processing (NLP) [A Complete Guide]. https://www.deeplearning.ai/resources/natural-language-processing/ 

Edgren, Gustaf, et al. (2007). Socioeconomic status, family history of cancer, and risk of Hodgkin Lymphoma.” Journal of the National Cancer Institute, 99(19), 2007, pp. 1476–1484. https://news.ki.se/ai-can-identify-women-with-high-risk-of-breast-cancer-in-screening-examinations 

Gal, E. (2021). Introducing the Ezra full-body scan, covering 13 cancers in women and 11 in men. Medium. https://medium.com/ezraai/introducing-the-ezra-full-body-scan-covering-13-cancers-in-women-and-11-in-men-67ba90b4e534.

Genetic Information Nondiscrimination Act (GINA). (2008). Retrieved from https://www.hhs.gov/hipaa/for-professionals/special-topics/genetic-information/index.html 

Hamilton, J.G., Genoff, G.M., Westerman, J.S.,  Shuk, E., Hay, J.L., Walters, C., Elkin, E., Bertelsen, C., Cho, J., Daly, B., Gucalp, A., Seidman, A.D., Zauderer, M.G., Epstein, A.S., & Kris, M.G. (2019). “A Tool, Not a Crutch”: Patient Perspectives About IBM Watson for Oncology Trained by Memorial Sloan Kettering. J Oncol Pract,15(4),e277-e288. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6494242/ 

Janssens, A. C. J. W., & Martens, F. K. (2020). Reflection on modern methods: revisiting the area under the ROC Curve, International Journal of Epidemiology, 49(4), p.1397–1403. https://pubmed.ncbi.nlm.nih.gov/31967640/ 

LeCun, Y., Bengio, Y., Hinton, G. (2015). Deep learning. Nature 521,p 436–444. https://www.nature.com/articles/nature14539 

Malin, J. L. (2013). Envisioning Watson as a rapid-learning system for oncology. Journal of oncology practice, 9(3), 155–157. https://pubmed.ncbi.nlm.nih.gov/23942497/ 

Manolio, T. A., Collins, F. S., Cox, N. J., et al. (2009). Finding the missing heritability of complex diseases. Nature, 461(7265), 747-753. https://www.nature.com/articles/nature08494 

Marinovich, M.L., Wylie,E.,  Lotter,W., Lund, H.,  Waddell,A.,  Madeley,C.,  Pereira,G.,& Houssamia, N. (2023) Artificial intelligence (AI) for breast cancer screening: BreastScreen population-based cohort study of cancer detection. The Lancet eBioMedicine,90 (104498). https://www.thelancet.com/journals/ebiom/article/PIIS2352-3964(23)00063-4/fulltext 

Mavaddat, N., Michailidou, K., Dennis, J., Lush, M., Fachal, L., Lee, A., & Chatterjee, N. (2019). Polygenic risk scores for prediction of breast cancer and breast cancer subtypes. American Journal of Human Genetics, 104(1), 21-34. https://www.cell.com/ajhg/fulltext/S0002-9297(18)30405-1

Microsoft UK Stories. (2024). AI detects 12% more breast cancers in UK first. https://ukstories.microsoft.com/features/12-more-breast-cancers-detected-with-potential-workload-savings-of-30-in-the-uks-first-prospective-evaluation-of-breast-screening-ai/#:~:text=An%20 artificial%20intelligence%20(AI)%20breast

Narkhede, S. (2022) Understanding AUC – ROC Curve – towards data science, Medium. https://towardsdatascience.com/understanding-auc-roc-curve-68b2303cc9c5

Park, T., Gu, P., Kim, C. H., Kim, K. T., et al. (2023). Artificial Intelligence in urologic oncology: the actual clinical practice results of IBM Watson For Oncology in South Korea. Science Direct, 11(4), 218-221. https://doi.org/10.1016/j.prnil.2023.09.001

Patel, S. (2024) AI-Powered Virtual Assistants Can Change Care Management. https://medtechintelligence.com/feature_article/ai-powered-virtual-assistants-can-change-care-management/ 

Patel, K., Huang, S., Rashid, A., Varghese, B., & Gholamrezanezhad, A. (2011). A narrative review of the use of artificial intelligence in breast, lung, and prostate cancer. Life 2023, 13(10):2011. https://www.mdpi.com/2075-1729/13/10/2011; https://www.mdpi.com/2075-1729/13/10/2011 

Pharaoh, P. D., Antoniou, A., Bobrow, M., Zimmern, R. L., Easton, D. F., & Ponder, B. A. (2002). Polygenic susceptibility to breast cancer and implications for prevention. Nature Genetics, 31(1), 33-36. https://pubmed.ncbi.nlm.nih.gov/11984562/

Predilife. (2024). MammoRisk.  https://www.predilife.com/en/mammorisk

Sarker, I.H. (2021) Machine learning: algorithms, real-world applications and research directions, SN COMPUT. SCI. 2, 160 https://doi.org/10.1007/s42979-021-00592-x

Shen, Y., Shamout, F.E., Oliver, J.R., Witowski, J., Kannan, K., Park, J., Wu, N., Huddleston, C., Wolfson, S., Millet, A., Ehrenpreis, R., Awal, Disha., Tyma, C., Samreen, N., Gao, Y., Chhor, C., Gandhi, S., Lee, C., Subaiya, S.K., Leonard, C., Mohammed, R., Moczulski, C., Altabet, J.,  Babb, J., & Geras, K.J. (2021). Artificial intelligence system reduces false-positive findings in the interpretation of breast ultrasound exams. Nature Communications, 12(1). https://www.nature.com/articles/s41467-021-26023-2