Cancer outcomes have been showing improvements in recent years, studies showing that fewer people between the ages of 35 and 69 are dying from the disease than at any point in the last 25 years. Advancements in artificial intelligence (AI), particularly in the subspace of medical image screening using AI, hold tremendous promise in the fight against cancer by improving accuracy of diagnosis and avoiding unnecessary medical interventions through predictive capabilities. However, the full potential of AI screening hinges on overcoming the challenge of ensuring access to high-quality, diverse data for accurate and equitable outcomes.
MIA, or mammography intelligent assessment is an AI platform for breast screening developed by Kheiron Medical Technologies and Imperial College London. A study published in Nature Medicine found that MIA could increase early detection of breast cancers in a European health setting by up to 13% in comparison to the clinical gold-standard of analysing every case by two different radiologists. Of the additional cancers found, 83% were invasive showing that MIA can detect cancers where early detection is vital. In a separate UK test MIA correctly detected all the scans identified by NHS clinicians as showing signs of breast cancer in set of 10,000 mammograms. In addition, MIA correctly detected 11 scans that the clinicians had missed.
In a study at the Department of Pathology and Immunology at Washington University School of Medicine, researchers trained a machine learning algorithm to predict the onset of brain cancer in patients with early-stage lung cancer using biopsy samples from patients. The algorithm was able to predict the development of brain cancer with 87% accuracy in comparison to 57.3% accuracy achieved by a team of four pathologists. A small percentage of early-stage lung cancer patients will eventually develop brain cancer, so the improved ability to predict the development dispenses with the need for aggressive intervention out of caution.
Challenges to improvements in the accuracy of AI screening include good quality datasets which form the reservoir of information that the AI algorithm ‘learns’ in order to assess subsequent new images, for example, in real time. Good datasets can be progressively more difficult to come by for rare conditions and underrepresented groups of people.
A review in the International Journal of Dermatology identified significant shortcomings when AI was applied to non-white pigmentations, or ‘skin of colour’. A UK government commissioned review into unfair biases in the design and use of medical devices found that image screening using AI suffers from an inherent bias against certain groups in the population, notably women, people from ethnic minorities and disadvantaged socio-economic groups. One example is the potential under-diagnosis of skin cancer for people with darker skin tones, a possible consequence of AI systems trained predominantly on images of lighter skin tone.
AI systems for reading chest x-rays are mainly trained on images of men, who tend to have larger lung capacity, potentially worsening the likelihood of accurate diagnosis for women. Given a dataset in which women are underdiagnosed for a condition could lead to an AI model associating being female with being at lower risk for that condition.
While the UK government says it is committed to removing bias in datasets and increasing training for healthcare professionals, good quality data sets remain an Achilles heel for AI in the screening of medical images and the ‘black box’ nature of an AI algorithm leaves its real world accuracy at the mercy of how well its ‘learned’ dataset reflects the real world.
References: https://eandt.theiet.org/2024/07/09/ai-winning-battle-against-cancer