Reshaping Cancer Diagnosis: AI-Powered Detection

Authored By: Manuel Rodríguez Villegas

Art By: Amy Em

Early detection of carcinogenic cells is crucial for improving patient outcomes. In fact, identifying cancer at Stage 1 significantly increases the survival probability compared to Stage 2 [1]. However, milder symptoms, potentially unnecessary tests, and likely-but-incorrect diagnostics make it challenging for radiologists to detect the problem in advance. At this stage, the introduction of Artificial Intelligence (AI) classification systems could be highly beneficial, offering non-trivial predictions that the human eye may miss. As will be discussed, AI-assisted diagnosis is already improving patients’ outcomes in different fields. However, the rise of this technology has also awakened strong criticism and calls for action.

The term AI, so widely used nowadays, refers to algorithms that enable computers to carry out tasks typically requiring human cognition. These algorithms process vast amounts of data and gradually adjust their internal parameters to improve their predictions. When dealing with images, the most common approach is to use Convolutional Neural Networks (CNNs), which offer useful properties such as translation invariance. Recurrent Neural Networks (RNNs) and Transformers, the system behind LLMs, have also been widely applied in different medical use cases. The flexibility of these models, alongside the increased computational power and data availability, is one of the key factors that has contributed to the success of this technology during the last few years.

One of the earliest applications of AI to healthcare was MYCIN [2], a program developed during the 1970s at Stanford University, which could provide advice and explanations regarding the diagnosis of bacterial infections. Since then, state-of-the-art AI has seen major developments. A notable example is Mirai [3], a model developed by MIT that was recently evaluated in the most extensive study of an AI-based breast cancer model to date. The experiment included data from hospitals in five different countries and focused on comparing guidelines for selecting patients for additional screening MRI. The comparison was done between guidelines based on both Mirai and Tyrer-Cuzick (TR) lifetime risk, a widely adopted risk model that also analyzes family history and patient demographics. The study examined whether early detection could be improved without increasing costs (keeping the true negative rate the same), or if costs could be lowered while still catching the same number of real cases (keeping the true positive rate the same). The results demonstrated the superiority of Mirai’s guidelines, achieving a 70% improvement in sensitivity compared to TC at the same specificity. Moreover, Mirai’s performance remained consistent across all test sets and did not exhibit common biases that traditional methods often display against minorities, usually due to underrepresentation in training data and overfitted models.

The Mirai case is just one of many recent successes. The Transformer architecture has also been leveraged to develop a scalable pancreatic cancer recognition system [4]. This type of cancer is not very common but presents an elevated mortality rate due to the delayed onset of symptoms. Since large-scale population screening is not a feasible option, the study proposes using clinical records to identify high risk patients, which could participate in specific surveillance programs. Experimental results display how the Transformer and the GRU models achieve very promising results both in AUROC and relative risk (RR) metrics.

Despite the huge potential of these algorithms, there are still medical subfields in which AI needs time to develop. A clear example is lung cancer computed tomography (CT) screening. Lung cancer is the leading cause of cancer-related fatalities worldwide, which makes it a strong candidate for AI-assisted diagnosis. It is visually manifested on CT scans as opacities in the lung parenchyma which are not part of the original anatomy, commonly known as pulmonary nodules. Nodule detection suffers from a high rate of false positives, but no AI system has been capable of outperforming radiologist’s accuracy when predicting them [5]. The most advanced algorithms perform on par with professionals, while those claiming superior results often lack sufficient detail to be considered credible. However, combining experts and machines results could lead to more robust performance as each agent mitigates the other’s mistakes.

The rise of artificial intelligence has also induced various ethical concerns. Bias in training data has previously led to significant drops in accuracy for underrepresented populations [6], promoting distrust and skepticism towards AI-generated results. Furthermore, privacy concerns emerge when training methods of huge technological companies are analyzed [7]. Policymakers should require explicit user consent before allowing AI companies to train models with personal data. Responsibility is another crucial issue, especially in cases where AI failures could have serious consequences. AI engineers, in collaboration with regulatory agencies, should guarantee that their products align with fundamental human rights (privacy, non-discrimination, freedom of speech, etc.) as they become part of our daily lives.

AI is no longer a distant technology, and its applications are starting to demonstrate incredible results in many different sectors. Cancer detection has shown a huge potential for improvement through advances in medical images recognition and clinical records classification. While significant progress is still needed, as well as a thorough analysis of ethical concerns, AI clearly has the potential to start a new era for healthcare.

Manuel Rodríguez Villegas

References

[1] Crosby, D. et al. (2022, March 18). Early detection of cancer. Science, 375(6586). https://doi.org/10.1126/science.aay9040

[2] van Melle, W. (1978, May). MYCIN: a knowledge-based consultation program for infectious disease diagnosis. International Journal of Man-Machine Studies, 10(3), 313-322. ScienceDirect. https://doi.org/10.1016/S0020-7373(78)80049-2

[3] Yala, A. et al. (2022, June 1). Multi-Institutional Validation of a Mammography-Based Breast Cancer Risk Model. Journal of Clinical Oncology, 40(16), 1732-1740. https://doi.org/10.1200/JCO.21.01337

[4] Placido, D. et al. (2023, May 8). A deep learning algorithm to predict risk of pancreatic cancer from disease trajectories. Nature Medicine, 29, 1113–1122. https://doi.org/10.1038/s41591-023-02332-5

[5] Schreuder, A. et al. (2021, May 28). Artificial intelligence for detection and characterization of pulmonary nodules in lung cancer CT screening: ready for practice? Translational Lung Cancer Research, 10. https://doi.org/10.21037/tlcr-2020-lcs-06

[6] Buolamwini, J., & Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of Machine Learning Research, 81, 77-91. https://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf

[7] Morrison, S. (2023, July 27). OpenAI, Google, and Meta used your data to build their AI systems. Vox. https://www.vox.com/technology/2023/7/27/23808499/ai-openai-google-meta-data-privacy-nope