Abstract:
Medical image analysis has significantly advanced with machine learning, enhancing disease diagnosis and biomaterial toxicity assessment. However, challenges such as data heterogeneity, computational complexity, and the lack of standardized evaluation metrics hinder its full potential. The study investigates supervised classification of three morphologically similar cell types, evaluating architectures (VGG16, Inception V3, SqueezeNet) and classifiers (Neural Network, Random Forest, KNN, etc.). Results indicate VGG16 paired with Neural Networks achieves the highest accuracy. Unsupervised clustering is explored by applying ISO guidelines to assess biomaterial toxicity levels, leveraging VGG16 and SqueezeNet for feature extraction. A hybrid clustering approach enhances classification into toxicity levels, demonstrating improved separability with high-pass filtering techniques. A U-Net-based model is optimized for cell counting, evaluating optimizer-loss function combinations for segmentation and confluency analysis.
v
Experiments on cells exposed to biomaterials (PAR50, UniFast) reveal toxicity patterns through morphological changes. Hybrid loss functions (Dice-Focal, Jaccard-BCE) significantly improve segmentation accuracy. Quantization and pruning techniques are applied to reduce computational demands without compromising accuracy to enable real-world deployment. A pruned U-Net achieves 95% segmentation accuracy. This research contributes novel methodologies for biomedical image analysis by: (i) developing a benchmarked unsupervised clustering framework aligned with ISO standards, (ii) proposing a high-accuracy classification model for cell types, (iii) optimizing U-Net for segmentation and counting, and (iv) enhancing computational efficiency through model compression. These findings support automated biomaterial toxicity assessment, improving efficiency and standardization in medical imaging applications.
