Work

Transfer learning for multi-class oral epithelial dysplasia grading: architecture, activation, and split-size effects on training convergence and test performance

Year: 2026
Type: article
Abstract: Oral epithelial dysplasia (OED) is a critical precursor to oral squamous cell carcinoma, making accurate grading vital for timely clinical intervention. This study develops an automated patch-based deep learning framework to improve multi-class OED grading accuracy from histopathological tissue images. Using a locally acquired dataset of 3,777 tissue strips from 94 patients, 256 × 256 images were enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE) and split into four non-overlapping 128 × 128 patches, generating 15,108 training samples. Seven transfer learning architectures (VGG16, ResNet152V2, InceptionV3, MobileNetV2, Xception, DenseNet201, and InceptionResNetV2) were systematically evaluated using two activation functions (ReLU, Leaky ReLU) across 80:20 and 90:10 data splits. Training convergence, validation performance, and early stopping behavior were tracked across all configurations. Xception with ReLU activation under a 90:10 split achieved the highest held-out test performance, reaching 85.19% classification accuracy and a weighted F1-score of 85.36%, along with top validation accuracy (84.77%). Conversely, VGG16 with Leaky ReLU and InceptionResNetV2 with ReLU (both 90:10 split) suffered complete training collapse. The proposed patch-based transfer learning framework demonstrates strong potential for reliable, automated multi-class OED grading. Additionally, instances of model collapse highlight important optimization instabilities to navigate when fine-tuning deep architectures on histopathological data.
Language: English
Open Access status: gold
Estimated APC paid: $2,490
OpenAlex ID: W7237206692
FWCI:
Cites: 28
Cited by:
Related to: 10