GhostTuDNet: accelerating tumor detection on CPUs with a compact and efficient model

Early detection of tumors, particularly intracranial and breast tumors, is essential for improving patient outcomes. Current deep learning-based algorithms have shown remarkable success in identifying tumors but often require high computational power, limiting their use in resource-constrained environments. To address this challenge, we developed GhostTuDNet, a compact and efficient model based on GhostNetV3, specifically optimized for CPU operation. The model was trained using knowledge distillation with ViG as the teacher model and was evaluated on two datasets: breast ultrasound and brain MRI. GhostTuDNet achieved a Precision of 93.5% and a Recall of 92.7% on the Intracranial Tumor dataset, closely matching the performance of the ViG model, which recorded a Precision of 94.2% and Recall of 93.4%. Despite having only 6.1M parameters and an inference time of 8.7ms, GhostTuDNet's performance was comparable to the much larger ViG model, which has 86.8M parameters and a 231ms inference time. In contrast, MobileNetV3, another compact model, showed significantly lower performance, with a Precision of 89.8% and Recall of 88.2% on the Intracranial Tumor dataset. These results highlight GhostTuDNet's ability to balance compactness and accuracy, making it a viable solution for tumor detection in environments where computational resources are limited. Future work will explore the impact of class rebalancing techniques and evaluate GhostTuDNet across a broader range of datasets and platforms.