Parameter-Efficient Adaptation of Vision Transformers for Retinal OCT Classification
Students & Supervisors
Student Authors
Supervisors
Abstract
Vision Transformers (ViTs) deliver strong retinal optical coherence tomography (OCT) classification performance, but full fine-tuning is computationally heavy and can overfit limited medical datasets. We benchmark parameter-efficient fine tuning (PEFT) for a DeiT-Small ViT across five public OCT datasets (OCT2017, NEH-UT, Srinivasan2014, THOCT1800, and OCT-C8). We compare linear probing (head-only), VPT-Deep, AdaptFormer, and LoRA against standard full fine-tuning. PEFT methods consistently reduce trainable parameters to well under 3% while remaining competitive with full tuning. On the 8 class OCT-C8 benchmark, with frozen-backbone, AdaptFormer achieves the best Macro-F1 (97.6%), and LoRA reaches 96.9% while training only 0.7% of parameters, both outperforming full fine-tuning (95.6%). Across NEH-UT, Srinivasan, and THOCT, VPT-Deep/AdaptFormer/LoRA generally match or exceed full fine-tuning; OCT2017 is the main exception where full tuning remains superior. Calibration also holds up: PEFT variants show comparable or lower expected calibration error (ECE ⇡ 7–9%) than the fully tuned ViT. These results show that ViTs can be adapted to diverse OCT classification tasks with minimal weight updates, enabling cheaper training and storage and supporting deployment of multiple task-specific adapters on a shared backbone in compute and data-constrained clinical settings.
Keywords
Publication Details
- Type of Publication:
- Conference Name:
- Date of Conference: 05/07/2026 - 05/07/2026
- Venue: Dhaka Internation University