Classification (src/infer.py, src/model.py)

MobileViT Engine (src/model.py)

• Model: MobileViT-S (~5.6M parameters).
• Why: Combines the local inductive bias of CNNs with the global receptive field of Vision Transformers. This allows it to capture subtle chromatin textural changes in small regions at near-real-time speeds on edge devices.

Inference Loops (src/infer.py)

Prepares cell-level bounding boxes and feeds them directly through the MobileViT model.

Contains the logic for batch inferencing, evaluating precision/recall/F1-score across all 5 clinical categories, and appending raw softmax probabilities onto the Cell data contracts.

Categories Assessed

1. Superficial-Intermediate (Normal)
2. Parabasal (Normal)
3. Metaplastic (Benign/Reactive)
4. Koilocytotic (Low Grade Lesion)
5. Dyskeratotic (High Grade Lesion)