Can changing how a computer represents tissue colors improve its classification of melanoma samples?

Details

The researchers experimentally compared machine-classification approaches for melanoma biopsies that had been sectioned and stained with hematoxylin/eosin (H&E) on tissue microarrays (TMA). Seven tissue types represented four stages of melanoma progression: benign nevus; primary tumors with radial and vertical growth patterns (stage I); and four secondary metastatic tumor types—subcutaneous (stage II), lymph node (stage III), gastrointestinal and soft tissue (stage IV). The experimental setup comprised 14,208 image samples from 164 TMA cores.

The researchers constructed an HE color space by digitally deconvolving RGB images into separate H (hematoxylin) and E (eosin) channels. They compared three classifiers: Weighted Neighbor Distance (WND), Radial Basis Functions (RBF), and k-Nearest Neighbors (kNN).

The HE color space consistently outperformed other color spaces with all three classifiers, whereas differences between classifiers had a smaller effect on accuracy. The researchers interpreted this as showing that a more physiologically relevant representation of color can have a larger effect on correct image interpretation than downstream processing steps.

When the dataset was randomly split into training and test fields, individual fields of view were classified with an average accuracy of 96%. Testing entire cores not previously used in training produced 100% classification accuracy: four random trials used one test core for each of seven classes, for 28 tests in total. Each core corresponded to a different patient, so this test more closely mimicked a clinically relevant setting in which new patients are evaluated using training from previous cases.

The analysis method contained no parameters or adjustments specific to melanoma morphology. The researchers therefore suggested that it could be used to analyze other tissues and phenotypes, and potentially different image modalities and contrast techniques.