Combined Deep Learning Pipeline for Fully Automatic Classification and Segmentation of the Left Ventricle in Echocardiography

Abstract

Cardiovascular diseases remain a leading cause of death worldwide, and accurate automated analysis of echocardiographic images has a direct impact on early diagnosis and clinical decision making. In this thesis, we propose a deep learning framework for automatic multi-class cardiac classification and left ventricular segmentation from echocardiographic data, targeting three conditions: Normal, Heart Failure with Reduced Ejection Fraction (HFrEF), and Left Ventricular Hypertrophy or Hypertrophic Cardiomyopathy (LVH/HCM), which are also encountered in clinical practice in Algeria. Six architectures are evaluated for classification Swin-Tiny, ConvNeXt-Small, EfficientNet B4, ResNet-50, VGG-16, and a custom Classic CNN trained on a heterogeneous collection combining EchoNet-Dynamic, EchoNet-LVH, and CAMUS. For segmentation, four models are compared: UNet++ with a MiT-B3 transformer encoder, FPN, MANet, and a prompt-free MedSAM adaptation. Explainability is addressed through Grad-CAM++ and Attention Rollout visualizations. A clinical desktop application, LV-DeepSeg, is developed with an integrated LLM module generating structured medical reports via the Anthropic Claude API. Statistical validation includes Decision Curve Analysis, Precision-Recall curves, ROC analysis, and McNemar tests. The results demonstrate that Swin-Tiny achieves the best classification performance with a test accuracy of 92.36%, and UNet++ achieves an LV Cavity Dice of 0.9439 and an LV Wall Dice of 0.8784 with 97.46%-pixel accuracy, outperforming all other segmentation models. The findings confirm that competitive performance is achievable using ImageNet-pretrained backbones fine tuned on cardiac data, that transformer-based architectures generalize more reliably than convolutional networks under limited data, and that prompt-free SAM-based models fail to generalize to echocardiographic ultrasound without full encoder retraining.