Fetal Heart IQA — EfficientNetV2-S Quality Classifier

Theodore's Heart Initiative (THI) — Proof of Concept, Phase 3

EfficientNetV2-S trained to classify fetal cardiac ultrasound crops as GREEN (optimal), YELLOW (adjusting), or RED (searching/poor). Used as Step 2 of the two-step CHD screening pipeline:

YOLO11n (localization) → EfficientNetV2-S IQA (quality) → OpenCV overlay

Model Details


Architecture	EfficientNetV2-S, ImageNet pretrained backbone
Task	3-class image quality classification
Input	224×224 RGB crop of YOLO-detected fetal heart ROI
Output	Logits for [green, yellow, red] + view head stub
Training data	FOCUS dataset (300 expert 4CV images → 1,500 augmented crops)
Training platform	Kaggle T4 ×2, PyTorch 2.10.0+cu128, DataParallel
Epochs	75 per fold
Optimizer	AdamW (lr=3e-4, wd=1e-4) + CosineAnnealingLR

Training Results (5-Fold Cross-Validation)

Metric	Value
Mean val accuracy	99.89% ± 0.21%
Test accuracy	97.62%
GREEN test acc	88.10%
YELLOW test acc	100.00%
RED test acc	100.00%
Best fold	Fold 1

Files

File	Description
`best_iqa_fold1.pth`	Best fold (use for inference)
`best_iqa_fold{2-5}.pth`	Remaining folds for ensemble use
`iqa_training_summary.json`	Full training history and per-class metrics

Usage

import torch
import torch.nn as nn
from torchvision.models import efficientnet_v2_s
from torchvision import transforms
from PIL import Image

LABELS = ["green", "yellow", "red"]

class IQAModel(nn.Module):
    def __init__(self, num_quality_classes=3):
        super().__init__()
        base = efficientnet_v2_s(weights=None)
        in_features = base.classifier[1].in_features
        base.classifier = nn.Identity()
        self.backbone = base
        self.quality_head = nn.Sequential(
            nn.Dropout(p=0.3), nn.Linear(in_features, 256),
            nn.SiLU(), nn.Dropout(p=0.2), nn.Linear(256, num_quality_classes),
        )
        self.view_head = nn.Sequential(nn.Dropout(p=0.3), nn.Linear(in_features, 1))

    def forward(self, x):
        f = self.backbone(x)
        return self.quality_head(f), self.view_head(f)

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

model = IQAModel()
model.load_state_dict(torch.load("best_iqa_fold1.pth", map_location="cpu"))
model.eval()

img = Image.open("your_crop.png").convert("RGB")
with torch.no_grad():
    q_logits, _ = model(transform(img).unsqueeze(0))
label = LABELS[q_logits.argmax(1).item()]
print(label)  # "green", "yellow", or "red"

Limitations & Scope

Trained on 4-Chamber View (4CV) only — FOCUS dataset (single-center, 300 originals).
YELLOW and RED classes are synthetically generated from GREEN images via augmentation (blur, rotation, contrast, occlusion, noise). Real-world YELLOW/RED examples were not available at POC stage.
Not validated on LVOT, RVOT, or 3VV views.
Not for clinical use.

Citation

Theodore's Heart Initiative, 2026. Internal POC — fetal CHD screening research.

Downloads last month: -; Downloads are not tracked for this model. How to track

cellerson11
/

fetal-heart-iqa