Advancing Deep Learning in Medical Imaging Diagnostics under Small-Scale Dataset Constraints

Deep learning has brought clear gains to medical image analysis, but most clinical datasets remain small and heterogeneous, which often leads to overfitting, unstable training, and unclear design choices for the model architecture and training strategy. This thesis focuses on small-scale dataset constraints in medical imaging diagnostics and proposes both a theoretical view and three concrete methods for multiple sclerosis (MS) and brain tumour classification. First, we introduce the Modelling Small-Scale Dataset Challenges Framework (MSSDCF), which links three core factors—data scale, model capacity, and optimisation behaviour—and explains why performance degrades when these factors are not matched. MSSDCF also offers three practical perspectives (sample, network, and search strategy) that guide the design of later methods. From the sample perspective, we propose Texture Embedded Multi-Stage Transfer Learning (TE-MSTL) for MS MRI classification. TE-MSTL embeds Gray-Level Co-occurrence Matrix (GLCM) texture images into a three-stage transfer pipeline. On a public MS MRI dataset, TE-MSTL improves the 10-fold cross-validated accuracy of a 6-layer ConvNet from 95.82% to 97.46% and raises the Matthews correlation coefficient from 91.65% to 93.02%. From the network perspective, we design RanMerFormer, a Vision Transformer with token merging and a random vector functional link head for brain tumour MRI. On a four-class public brain tumour dataset, RanMerFormer increases accuracy from 98.90% (ResNet152V2 / MobileNet + M-SVM) to 99.80%, while reducing FLOPs to about 70% of ResNet152V2 (8.2 G vs. 11.6 G) and remaining feasible on a single GPU. From the search strategy perspective, we propose PDAC-BTNet, a plane-decoupled adaptive curriculum network that combines MobileNetV2 with plane-specific curricula. On the same brain tumour dataset, PDAC-BTNet reaches 99.39% accuracy and 99.80% specificity with only 3.5 M parameters and 0.59 G FLOPs, and improves accuracy over RanMerFormer by 0.52 and 1.32 percentage points in two simulated imbalance scenarios. Together, these results show that the MSSDCF-guided designs can give consistent and quantifiable gains on small-scale brain MRI benchmarks while keeping model and computational costs under control.