CoLoRA: Efficient Fine-Tuning for CNNs with OCT Image Classification

CoLoRA (Convolutional Low-Rank Adaptation) has been unveiled by researchers as an efficient method for fine-tuning convolutional neural networks (CNNs) with minimal parameters. This approach builds upon the LoRA technique, breaking down kernel updates into lightweight depthwise and pointwise elements. By doing so, it decreases the number of trainable parameters for convolutional updates by over 80% compared to traditional fine-tuning, while enabling the integration of these updates into existing pretrained convolutional kernels, thus maintaining the model's original size and inference complexity. Tests on MedMNIST datasets, especially OCTMNISTv2, reveal that CoLoRA, when applied to VGG16 and ResNet50, achieves competitive classification results with significantly fewer trainable parameters. The method's trade-offs in predictive performance and parameter efficiency are explored through comparisons with transfer learning, adapters, BitFit, and convolutional LoRA variants. Detailed in a paper on arXiv (2505.18315), this work, announced as a replace-cross on an unspecified date, is significant for medical image analysis, particularly in optical coherence tomography (OCT) image classification, providing a practical solution for fine-tuning large convolutional models under limited computational resources.