When Fine Tuning Does Not Help: A Two Stage EfficientNet B0 Evaluation for Four Class Lung CT Classification on a Small Cohort

This study evaluates standard transfer learning practices on small medical datasets using an EfficientNet-B0 model trained on a 990-image chest CT collection. The authors’ primary contributions demonstrate that partial fine-tuning (unfreezing the top backbone layers) yielded no improvement over frozen feature extraction, reaching its peak validation accuracy (88.89%) prior to fine-tuning. Additionally, the paper highlights a significant divergence between metrics—where a high macro ROC-AUC (0.9466) masked lower test accuracy (76.83%) and poor recall (0.58) for adenocarcinoma—showing that ROC-AUC alone misrepresents practical model quality. Finally, the study illustrates how balanced class weighting merely redistributed errors by over-predicting minority classes, and it documents a critical label space mismatch in the source dataset requiring explicit harmonization.