Feature Imitating Networks Enhance The Performance, Reliability And Speed Of Deep Learning On Biomedical Image Processing Tasks

TL;DR

This paper introduces Feature Imitating Networks (FINs), pre-trained to approximate closed-form radiomics features, and embeds them into larger neural networks to enhance biomedical image processing tasks. By training six FINs to imitate six radiomics features and integrating them into models for COVID-19 CT detection, brain MRI classification, and brain-tumor segmentation, the authors demonstrate improved accuracy, faster convergence, and reduced training variance compared with baselines of similar capacity. Importantly, FINs do not simply provide raw feature values; they adapt the feature representations during task-specific fine-tuning to achieve task-relevant representations. The results suggest FINs can deliver state-of-the-art performance across biomedical imaging tasks, particularly in data-scarce settings, with faster, more reliable training dynamics and potential for broader applicability across domains.

Abstract

Feature-Imitating-Networks (FINs) are neural networks that are first trained to approximate closed-form statistical features (e.g. Entropy), and then embedded into other networks to enhance their performance. In this work, we perform the first evaluation of FINs for biomedical image processing tasks. We begin by training a set of FINs to imitate six common radiomics features, and then compare the performance of larger networks (with and without embedding the FINs) for three experimental tasks: COVID-19 detection from CT scans, brain tumor classification from MRI scans, and brain-tumor segmentation from MRI scans. We found that models embedded with FINs provided enhanced performance for all three tasks when compared to baseline networks without FINs, even when those baseline networks had more parameters. Additionally, we found that models embedded with FINs converged faster and more consistently compared to baseline networks with similar or greater representational capacity. The results of our experiments provide evidence that FINs may offer state-of-the-art performance for a variety of other biomedical image processing tasks.

Figures (2)

• Figure 1: An illustration of our approach that integrates Feature Imitating Networks (FINs) into larger model structures for biomedical image processing tasks. A biomedical image (bottom) is passed to a CNN (blue), and a set of FINs (green), pre-trained to approximate a set of closed-form radiomics features (see Section \ref{['sec:Meth']}). The results from FINs and the CNN layers are received by a fully connected DFNN (orange) to predict the outcome. The representations learned by the FINs evolve during fine-tuning for the task.
• Figure 2: UNet training performance before (left column) and after (right column) insertion of FINs. Plots show loss (row 1), IoU (row 2), and dice coefficient (row 3) during the $50$ epochs of training: (a) loss of basic UNet,(b) loss of UNet with FINs, (c) IoU of basic UNet, (d) IoU of UNet with FINs, (e) dice coefficient of basic UNet and (f) dice coefficient of UNet with FINs. The red line represents the validation set loss metrics while the blue line represents for training set loss metrics. Overall training with FINs is observed to be more stable.