SkullFractureNet: A VGG16-Based Transfer Learning Pipeline with Two-Phase Fine-Tuning and Test-Time Augmentation for Automated Skull Fracture Detection in Head CT

Abstract

This work seeks to tackle the problem of significant misdiagnosis rates in CT imaging for identifying skull fractures – as high as 14.8%, even amongst expert radiologists – thus highlighting the need for reliable automatic skull fracture detection approaches. In this study, a new SkullFractureNet approach is proposed where VGG16 is used as the convolutional neural network (CNN) backbone, pretrained on the ImageNet database, instead of the previous efficient net based approach (EfficientNetB0). The proposed approach utilizes specific architectural modifications aimed at addressing overfitting issues arising from the small number of data samples in a real-world setting: GlobalMaxPooling2D is used in place of the GlobalAveragePooling2D layer to better locate local fracture areas; the classification head size is reduced from 256 to 64; dropout rate is increased to 0.6 and 0.4, and L2 regularization is applied throughout the dense layers. In terms of the training process, the same two-phase process remains – the first phase trains the classifier head while freezing all of the VGG16 weights (learning rate = 5 × 10⁻⁵; 40 epochs); the second phase unfreezes only the final convolutional block (block5_conv3) at a learning rate of 5 × 10⁻⁶. Aggressive slice-level augmentation strategies (flip, rotate, zoom, brightness, contrast, noise and cutout) in combination with a test-time augmentation (TTA) approach where eight forward passes are averaged together are used.The experiments were performed on a balanced dataset of 42 samples (21 fractures vs. 21 normals) using a patient-based split approach. The results of testing on the held-out portion of the test set (6 samples) yield the following performance metrics (maximum probability aggregation): accuracy 0.667, specificity 1.000, precision 1.000, AUC-ROC 0.667, and AUC-PRC 0.756, and no false positives. It can be noted that the training process shows that the measures adopted to prevent overfitting were effective.