Automated acute skin toxicity scoring in a mouse model through deep learning

Area of Science:

• Preclinical radiotherapy research
• Medical imaging analysis
• Artificial intelligence in oncology

Background:

• Skin toxicity is a common and undesirable side effect in radiotherapy.
• Manual scoring of skin reactions in preclinical trials can be subjective and time-consuming.
• Objective and reproducible assessment methods are crucial for evaluating radiotherapy efficacy and safety.

Purpose of the Study:

• To develop and validate a novel imaging and deep learning framework for objective skin toxicity assessment in preclinical radiotherapy.
• To compare the performance of the deep learning model against expert observers in evaluating toxicity.
• To address limitations of manual scoring, including inter-observer variability and evaluation time.

Main Methods:

• A dataset of 7542 images from 160 mice in proton/electron radiotherapy trials was created.
• A two-step deep learning model was developed, including object detection for hind leg identification and classification for toxicity grading.
• An observer study compared the deep learning model's performance with five expert scorers on retrospective data.

Main Results:

• The hind leg object detection model achieved nearly 99% accuracy.
• The toxicity classification model demonstrated an overall accuracy of approximately 85%.
• The deep learning model showed superior accuracy and reduced misclassification distance compared to human observers, with high inter-observer agreement.

Conclusions:

• The developed imaging and deep learning system offers an objective and reproducible method for skin toxicity assessment in preclinical radiotherapy.
• This system has the potential to minimize inter-observer variation and reduce evaluation times in manual scoring.
• Further refinement with larger datasets could enable deployment in preclinical research and clinical radiotherapy trials.