Deep learning (DL) to develop a triage tool for pediatric skin tumors

This project aims to leverage deep learning (DL) to develop a triage tool for pediatric skin tumors, enhancing diagnostic accuracy and streamlining access to specialist care.

Objectives

Primary Objective: Develop a DL-based tool to provide ranked differential diagnoses and triage recommendations for pediatric skin tumors.

Secondary Objective: Evaluate the performance of the algorithm with a prospective dataset and evaluate the tool’s clinical utility.

Research Plan

Aim 1: Algorithm Development and Validation

1. Retrospective Data Collection: Build a retrospective image dataset of single-lesion cutaneous tumors or mimickers from CHU Sainte-Justine’s medical photography archive and public sources.
2. Model Training: Train a convolutional neural network through transfer learning from pretrained ResNet34 models and utilize data augmentation techniques to enhance robustness, and to mitigate biases in rare conditions and darker skin types.
3. Validation: Evaluate the triage performance of the algorithm on the testing dataset by analyzing the algorithm’s impact on reducing unnecessary referrals and improving the identification of cases requiring dermatology consultations.

Aim 2: Prospective Validation and Clinical Usability

1. Prospective Data Collection: Build a prospective image dataset of 500 pediatric patients with cutaneous tumors over two years at CHU Sainte-Justine.
2. Prospective Validation: Assess algorithm performance on the prospective dataset using metrics such as diagnostic accuracy, sensitivity, specificity, and triage efficacy.
3. Reader Study: Conduct a study with 10 PCPs and 10 pediatricians to compare diagnostic accuracy, triage efficiency, confidence, and time per case with and without assistance of theAI algorithm.