Svetozar1993/LocalMedScan-malaria-mobilenetv2

MobileNetV2 (Fine-Tuned) — Malaria Parasite Detection

Recommended production model. Fine-tuned MobileNetV2 for malaria parasite detection from thin blood smear microscopy images. Achieves 97.97% test accuracy — the highest in our model family. Part of LocalMedScan.

Model Details

Property	Value
Architecture	MobileNetV2 (ImageNet pretrained, fine-tuned)
Parameters	~3.4M
Model Size	14 MB
Input	RGB 224x224 blood smear cell images
Output	Binary classification: Parasitized vs Uninfected
Training	3-phase progressive unfreezing from ImageNet weights
Framework	PyTorch
License	MIT

Performance

Test Results (NIH Malaria Dataset, 2,757 test images)

Metric	Without TTA	With 5-View TTA
Accuracy	97.68%	97.97%
Sensitivity	97.12%	97.41%
Specificity	98.22%	98.51%
Precision	98.19%	98.49%
F1 Score	97.65%	97.95%

Comparison with PlasmoSENet

Model	Params	Size	Accuracy (TTA)	F1	Training
MobileNetV2 (this model)	3.4M	14 MB	97.97%	97.95%	Fine-tuned from ImageNet
PlasmoSENet	10.6M	41 MB	96.55%	96.47%	Trained from scratch

MobileNetV2 outperforms PlasmoSENet by 1.42 percentage points while being 3x smaller and 3x faster. Transfer learning from ImageNet provides a decisive advantage on this dataset size (27K images).

Use MobileNetV2 for deployment. Use PlasmoSENet for research into domain-specific architectures or when ImageNet-free training is required.

Key Achievements

1. 97.97% test accuracy with 5-view TTA — exceeds most published results on this benchmark
2. 3-phase progressive unfreezing prevents catastrophic forgetting while maximizing adaptation
3. TransformSubset bug fix — corrected a common PyTorch random_split bug where train/val/test subsets inadvertently share transforms, causing training to run without augmentation
4. 5-view TTA exploiting blood cell rotational invariance (+0.29% accuracy)
5. Lightweight — 14 MB model runs on any device including mobile

Training Details

3-Phase Progressive Unfreezing

• Phase 1: Classifier head only (frozen backbone)
• Phase 2: Last inverted residual blocks unfrozen, reduced LR
• Phase 3: Full end-to-end fine-tuning, further reduced LR

Augmentation

Strong augmentation pipeline with RandomResizedCrop, RandomRotation(90), ColorJitter, GaussianBlur, and RandomErasing. The TransformSubset wrapper ensures training and validation use independent transforms.

Dataset

NIH Malaria Cell Images Dataset (Rajaraman et al., 2018)

• 27,558 annotated cell images from Giemsa-stained thin blood smears
• 13,779 Parasitized + 13,779 Uninfected (balanced)
• Split: 80% train / 10% val / 10% test (seed=42)

Usage

import torch
import torchvision.models as models

# Load model
model = models.mobilenet_v2(weights=None)
model.classifier[1] = torch.nn.Linear(model.last_channel, 2)
state_dict = torch.load("model.pth", map_location="cpu", weights_only=True)
model.load_state_dict(state_dict)
model.eval()

# Inference
from torchvision import transforms
from PIL import Image

transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

img = Image.open("blood_smear_cell.png").convert("RGB")
tensor = transform(img).unsqueeze(0)

with torch.inference_mode():
    probs = torch.softmax(model(tensor), dim=1)
    # probs[0][0] = Parasitized, probs[0][1] = Uninfected

Links

• GitHub: Svetozar-Technologies/LocalMedScan
• PlasmoSENet Model: Svetozar1993/LocalMedScan-malaria-plasmosenet
• Dataset: NIH Malaria Cell Images