Evaluating the Limits of a Mitosis Detection Model on Low-Power Hardware
How can you improve runtime of a mitosis detector while not sacrificing performance?
Mitotic figures are prognostic indicators in histopathology, but patch-based detection can require substantial central processing unit(CPU) inference time. This thesis evaluated the speed–quality trade-off for a Fully Convolutional One-Stage (FCOS) detector with a ResNet-18 backbone on MIDOG++.
Methods included runtime backends, structured pruning, recovery fine-tuning, INT8 quantization and confidence-threshold calibration on a split separate from the test set. Workstation execution was restricted to four physical cores, with separate Raspberry Pi 5 runtime diagnostics. On 111 images and 8,500 patches, the PyTorch FP32 baseline required 9,261.67 seconds end-to-end, with an F1-score of 0.83007. The Pruned60 model with OpenVINO INT8 required 1,829.27 seconds, with an F1-score of 0.81956, and achieved a 5.063× speedup and an absolute F1 decrease of 0.01051. INT8 provided the primary acceleration, and recovery experiments did not justify selecting a trained checkpoint over the unrecovered state. On the Raspberry Pi 5, Pruned60 ONNX Runtime INT8 required a median of 728.18 seconds end-to-end for three images and 228 patches, a 4.274× speedup over FCOS_18 PyTorchFP32 on the same workload. This diagnostic establishes runtime behaviour, not device-specific detection quality. Since pruning selection used part of the test split, independent validation remains necessary. Runtime gains depend on hardware and backend configuration, and the acceptability of the quality loss remains application-specific.