The Intelligent Total Body Scanner for Early Detection of Melanoma (iToBoS) tool is being developed to help clinicians make earlier, more personalised diagnoses of a particularly aggressive form of skin cancer, with funding from the European Union.
It works by feeding patient scans and other clinically relevant forms of data into a Cognitive Assistance framework that uses different AI models to generate melanoma development risk scores.
Clinicians need to be confident in the results they receive from tools like iToBoS. Otherwise, they risk misdiagnosing their patients; either by indicating that melanoma does not exist, when it does, or by indicating that melanoma does exist, when it doesn’t. These high-impact errors could also affect some patient groups disproportionately, based on attributes like gender or age.
The iToBoS counted on Trilateral Research to support the development of the tool due to their expertise in identifying and removing bias in AI systems and to help empower clinicians with high-quality decision-making. By analysing the iToBoS clinical risk assessment model, developed by the National Technical University of Athens (NTUA), to test how fairly it worked for male and female patients. To assess both the model’s performance and fairness, the overall task was approached as a binary classification, despite the final output being a melanoma risk estimation. Given this, we found that in cases where the tool predictions indicated melanoma development, the results were fairly equal across men and women.
However, in cases where the predictions were incorrect, we saw significant bias. Female patients were 26% times more likely to be incorrectly classified as at risk for future melanoma development – a traumatic experience that could result in unnecessary treatments and other harmful outcomes.
To correct this bias against female patients, we needed to continually adjust our training data before re-training and testing the clinical risk assessment model. Doing so posed further challenges. For instance, re-training the model resulted in the model’s overall precision score decreasing by 8%. Nevertheless, we were able to redress this balance successfully, and ensure the tool delivers an equal number of false positive rates for both male and female patients. Another issue that we needed to navigate was the relatively small size of the dataset. We only had access to limited data, which became even more restricted when we categorised it by attributes, such as gender.
The potential for AI-enabled cancer detection tools is immeasurable, and iToBoS will provide future developers with the framework needed to build one responsibly. We’re so proud of the role we’ve played, having helped to reduce the chances of a melanoma misdiagnosis affected by factors like gender and age.