Diagnosis Of Takotsubo Syndrome By Robust Feature Selection From The Complex Latent Space Of DL-based Segmentation Network

TL;DR

This work tackles the challenge of distinguishing Takotsubo Syndrome (TTS) from STEMI using echocardiogram videos by leveraging the latent space of a left-ventricle segmentation network. It introduces two robust feature selection methods—GradCAM kernel ranking (FSR) and LASSO-based selection (FSL)—to extract disease-relevant latent features from the LV encoder. Among classifiers evaluated (SVMC, MLP, RFC), the LASSO-based approach combined with RFC achieves the best results, reaching $0.82$ accuracy and $0.82$ F1, surpassing prior state-of-the-art, and demonstrating the value of segmentation-derived latent features for robust, interpretable diagnosis. The methodology offers a path toward reduced computational complexity and improved generalization in noisy echo data, with potential for short- and long-term prognosis.

Abstract

Researchers have shown significant correlations among segmented objects in various medical imaging modalities and disease related pathologies. Several studies showed that using hand crafted features for disease prediction neglects the immense possibility to use latent features from deep learning (DL) models which may reduce the overall accuracy of differential diagnosis. However, directly using classification or segmentation models on medical to learn latent features opt out robust feature selection and may lead to overfitting. To fill this gap, we propose a novel feature selection technique using the latent space of a segmentation model that can aid diagnosis. We evaluated our method in differentiating a rare cardiac disease: Takotsubo Syndrome (TTS) from the ST elevation myocardial infarction (STEMI) using echocardiogram videos (echo). TTS can mimic clinical features of STEMI in echo and extremely hard to distinguish. Our approach shows promising results in differential diagnosis of TTS with 82% diagnosis accuracy beating the previous state-of-the-art (SOTA) approach. Moreover, the robust feature selection technique using LASSO algorithm shows great potential in reducing the redundant features and creates a robust pipeline for short- and long-term disease prognoses in the downstream analysis.

Figures (3)

• Figure 1: Method workflow diagram. A U-Net shaped architecture is used for LV segmentation from echo. Note that, the echo dataset is a video dataset (2D+t). $\mathcal{I}$ is the input video and $\mathcal{M}$ is the corresponding segmentation mask. Here, only a frame is shown for visual convenience. Latent features from the final convolutional layer of the encoder are used for feature selection. Feature selection block is detailed in \ref{['Ft_reduction']}. The selected features from feature selection step are used to train binary classifier for disease prediction.
• Figure 2: Workflow of the feature selection using GradCAM. Bottleneck kernel weights are obtained through back-propagation. Then the Kernel features are selected based on their weighted ranks and frequency. Finally the reduced feature kernels are flattened and trained with binary learning classifier.
• Figure 3: a) Randomly sampled frames of the echo dataset. b) Accumulated GradCAM visualization of $32$ feature kernels of the corresponding echos of a) extracted from the bottleneck. c) Accumulated GradCAM visualization of $3$ feature kernels after selecting feature kernels using weighted ranking. d) Accumulated GradCAM visualization of the feature kernels after selecting features using LASSO. Note that, even with an approximately $90\%$ reduction of the features for c)-d), the highlighted regions are almost identical with b).