Lymphoma triage from H&E using AI for improved clinical management

Area of Science:

• Digital pathology
• Computational pathology
• Oncology

Background:

• Routine lymphoma diagnosis involves manual triage of lymph node biopsies, leading to delays and over-referral.
• Current triage methods place a significant burden on specialized haematopathology services.

Purpose of the Study:

• To develop and evaluate an automated artificial intelligence (AI) system for lymph node biopsy triage.
• To improve the accuracy and speed of lymphoma case referral, reducing diagnostic delays and unnecessary referrals.

Main Methods:

• A retrospective dataset of H&E-stained lymph node whole slide images (WSI) was utilized from two hospitals.
• The dataset included follicular lymphoma, diffuse large B-cell lymphoma, classic Hodgkin's lymphoma, and reactive controls.
• A machine learning model was trained on 80% of the data, validated on 10%, and tested on the remaining 10%.

Main Results:

• The AI triage system achieved a multiclass accuracy of 0.828 and an overall accuracy of 0.932 in discriminating between reactive and malignant cases.
• The AI's lymphoma detection capability was comparable to two haematopathologists and superior to a non-specialist pathologist.
• The AI tool incorporates uncertainty estimation and attention heatmaps to enhance explainability.

Conclusions:

• AI-powered automated triage shows significant potential for accurate and timely lymphoma diagnosis.
• This technology can enhance patient care and improve diagnostic workflow efficiency.
• AI tools can assist in reducing the burden on pathology services and optimize referral pathways.