Evaluating deep learning-based melanoma classification using immunohistochemistry and routine histology: A three...

Area of Science:

• Computational pathology
• Digital diagnostics
• Biomedical imaging analysis

Background:

• Pathologists utilize hematoxylin and eosin (H&E) and MelanA immunohistochemical (IHC) stains for accurate melanoma diagnosis.
• Deep learning (DL) systems excel in analyzing H&E slides but are less explored for IHC analysis.
• There is a need to evaluate DL performance on IHC slides for diagnostic support.

Purpose of the Study:

• To assess the diagnostic performance of DL models trained on MelanA IHC slides.
• To compare the performance of MelanA IHC DL models against H&E-based DL benchmarks.
• To investigate the combined performance of DL models using both MelanA IHC and H&E stains.

Main Methods:

• ResNet models were trained on MelanA IHC-stained and H&E-stained tissue slides.
• The performance of individual MelanA and H&E classifiers was evaluated on out-of-distribution (OOD) datasets.
• A combined classifier integrating both MelanA and H&E data was developed and tested.

Main Results:

• The MelanA DL classifier achieved an area under the receiver operating characteristics curve (AUROC) of 0.82 (OOD), comparable to the H&E benchmark (AUROC 0.81).
• The H&E benchmark classifier achieved an AUROC of 0.75 on OOD datasets, while the MelanA classifier achieved 0.74.
• The combined MelanA and H&E classifier demonstrated improved performance with AUROCs of 0.85 and 0.81 on OOD datasets.

Conclusions:

• DL-based assistance systems using MelanA IHC slides offer diagnostic performance similar to traditional H&E classification.
• Multi-stain DL classification, combining MelanA IHC and H&E, enhances diagnostic accuracy.
• These findings suggest potential for improved DL tools to aid pathologists in routine melanoma diagnosis.