Fig 1.
Pre-process PanNuke dataset [15] for nuclei instance segmentation and classification.
Fig 2.
An Overview of the proposed framework.
Different medical institutions locally train models on private data in two stages: segmentation followed by classification, with hyperparameter optimization applied to optimize the performance of both models. The optimized models are then quantized to reduce size and improve efficiency before being uploaded to a federated server. The server aggregates the quantized models, creating a robust, generalized model. Medical practitioners can download the aggregated model to improve healthcare insights while ensuring data privacy.
Fig 3.
An overview of the PanNuke nuclei distribution across the nineteen tissue types, sorted by the total number of nuclei within each tissue.
The total number of nuclei for each tissue type is provided in parentheses. Adapted from [27].
Table 1.
Search space for segmentation and classification models.
Table 2.
Quantitative evaluation of different segmentation models.
Table 3.
CPA and IoU comparison of different segmentation models.
Table 4.
Average mPQ and bPQ of the nuclei labels across the 19 different tissue types in the PanNuke dataset.
Table 5.
Average PQ of each nucleus category.
Table 6.
Specificity (SP), sensitivity (SN), F1-Score, Matthews correlation coefficient (MCC), and accuracy (ACC) for classification of nuclei types.
Fig 4.
Class-wise performance of the classification models for categorizing nucleus.
Fig 5.
Performance of the different combinations for the SegNet and the DenseNet121 models.
Fig 6.
Performance analysis of different quantization techniques on latency, energy consumption, and size.
Table 7.
Performance comparison of federated learning algorithms for nucleus segmentation with INT8 quantized models.
Fig 7.
Performance analysis of different FL algorithms on both models.
Table 8.
Compare the performance of the proposed framework with other state-of-the-art frameworks.