A Clinical Evaluation of AI Solutions Developed in the CHAIMELEON Project for Cancer: Prostate, Lung, Breast, Colon and Rectum

An in Silico External Clinical Validation of AI Solutions for Cancer Management in the CHAIMELEON Project. Applied to 4 Target Types of Cancer (Lung, Breast, Prostate and Colorectal), Collected Through the Routine Delivery of Health Care With no Enrolment Conditinos (Real World Data).

The goal of this observational study is to see how useful an experimental viewer and AI solutions are for clinicians in their daily work. The investigators want to find out if the AI helps clinicians interpret medical images for different types of cancer.

The AI solutions aim to:

• Classify whether prostate cancer is low or high risk
• Classify the histological subtype in breast cancer
• Estimate the life expectancy of patients with lung cancer
• Determine the size of colon cancer, lymph node involvement and the possibility of metastasis..
• Assess the invasion of sorrounding tissues in the case of rectum cancer. The study will involve clinicians from various centres who will review a set of cases not previously analysed by the AI. Clinicians will do this in two phases: first using only their own expertise and then with the help of the AI solutions.

The technical team want to see if the AI solutions assist clinicians and could become useful in the everyday clinical practice. Clinicians will complete a survey to share their feedback on the usability of the platform and how helpful the AI solutions are.

In order to conduct a robust clinical validation, the investigators have designed a study on the required sample size. The study is design to evaluate the role of an AI-assisted tool as a support for improving the daily clinical work. The investigators used an online website (https://statulator.com/SampleSize/ss2PP.html) for the calculation and use the "paired binary proportions" option. Using the case of prostate cancer, the investigators want to compare the probability of correct risk classification in prostate cancer by clinicians alone and/or guided by AI. The study will have a significance (α) = 0.05; power (β) = 80%; the analysis will be "two sided" and with equal group sizes.

An 10% improvement in cancer risk classification was observed when clinicians had access to an AI tool solution (Yilmaz et al.,). In addition, the authors reported that expert readers had an accuracy rate of 81% compared to 69% for novice readers when determining the Gleason score of lesions (a medical term used in pathology to classify the aggressiveness of cells in a tumour). The authors also assumed an 80% correlation between paired observations.

As a result, at least 60 new cases would be needed to evaluate the performance of the AI tool.

Inclusion Criteria:

• patients with an histological confirmation of cancer diagnosis (prostate, lung, breast, colon or rectum)
• availability of radiological images (MR for prostate and rectum, CT for lung and colon or mammographys for breast).
• enough follow up (12 months for prostate, breast and rectum), 18 months for lung, and 24 months for colon.

Exclusion Criteria:

• patients with incomplete or low quality data (radiological, pathological or uncomplete clinical data necessary for the ground truth)