Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Comorbidity-Informed Transfer Learning for Neuro-developmental Disorder Diagnosis

Xin Wen, Shijie Guo, Wenbo Ning, Rui Cao, Jie Xiang, Xiaobo Liu, Jintai Chen

TL;DR

This work tackles the challenge of diagnosing neuro-developmental disorders from fMRI by addressing data scarcity and the need for robust representations. It introduces Comorbidity-Informed Transfer Learning (CITL), which exploits the ASD–ADHD comorbidity to guide cross-diagnostic transfer, using a TransferNet to filter comorbidity patterns and a pseudo-labeling–driven Enhanced Representation Generator to produce discriminative, i.i.d.-compliant representations. A depthwise separable convolution classifier, trained with a dual objective (Cross-Entropy and Cosine Embedding Loss), and a Conversion Engine autoencoder refine the dynamic functional connectivity features into an Optimized FC that supports robust diagnosis. On ABIDE I and ADHD-200 datasets, CITL achieves state-of-the-art accuracies of 76.32% and 73.15%, respectively, demonstrating that comorbidity-informed transfer and targeted representation enhancement can significantly improve neuropsychiatric disorder detection and offer interpretable connectivity-based insights.

Abstract

Neuro-developmental disorders are manifested as dysfunctions in cognition, communication, behaviour and adaptability, and deep learning-based computer-aided diagnosis (CAD) can alleviate the increasingly strained healthcare resources on neuroimaging. However, neuroimaging such as fMRI contains complex spatio-temporal features, which makes the corresponding representations susceptible to a variety of distractions, thus leading to less effective in CAD. For the first time, we present a Comorbidity-Informed Transfer Learning(CITL) framework for diagnosing neuro-developmental disorders using fMRI. In CITL, a new reinforced representation generation network is proposed, which first combines transfer learning with pseudo-labelling to remove interfering patterns from the temporal domain of fMRI and generates new representations using encoder-decoder architecture. The new representations are then trained in an architecturally simple classification network to obtain CAD model. In particular, the framework fully considers the comorbidity mechanisms of neuro-developmental disorders and effectively integrates them with semi-supervised learning and transfer learning, providing new perspectives on interdisciplinary. Experimental results demonstrate that CITL achieves competitive accuracies of 76.32% and 73.15% for detecting autism spectrum disorder and attention deficit hyperactivity disorder, respectively, which outperforms existing related transfer learning work for 7.2% and 0.5% respectively.

Comorbidity-Informed Transfer Learning for Neuro-developmental Disorder Diagnosis

TL;DR

This work tackles the challenge of diagnosing neuro-developmental disorders from fMRI by addressing data scarcity and the need for robust representations. It introduces Comorbidity-Informed Transfer Learning (CITL), which exploits the ASD–ADHD comorbidity to guide cross-diagnostic transfer, using a TransferNet to filter comorbidity patterns and a pseudo-labeling–driven Enhanced Representation Generator to produce discriminative, i.i.d.-compliant representations. A depthwise separable convolution classifier, trained with a dual objective (Cross-Entropy and Cosine Embedding Loss), and a Conversion Engine autoencoder refine the dynamic functional connectivity features into an Optimized FC that supports robust diagnosis. On ABIDE I and ADHD-200 datasets, CITL achieves state-of-the-art accuracies of 76.32% and 73.15%, respectively, demonstrating that comorbidity-informed transfer and targeted representation enhancement can significantly improve neuropsychiatric disorder detection and offer interpretable connectivity-based insights.

Abstract

Neuro-developmental disorders are manifested as dysfunctions in cognition, communication, behaviour and adaptability, and deep learning-based computer-aided diagnosis (CAD) can alleviate the increasingly strained healthcare resources on neuroimaging. However, neuroimaging such as fMRI contains complex spatio-temporal features, which makes the corresponding representations susceptible to a variety of distractions, thus leading to less effective in CAD. For the first time, we present a Comorbidity-Informed Transfer Learning(CITL) framework for diagnosing neuro-developmental disorders using fMRI. In CITL, a new reinforced representation generation network is proposed, which first combines transfer learning with pseudo-labelling to remove interfering patterns from the temporal domain of fMRI and generates new representations using encoder-decoder architecture. The new representations are then trained in an architecturally simple classification network to obtain CAD model. In particular, the framework fully considers the comorbidity mechanisms of neuro-developmental disorders and effectively integrates them with semi-supervised learning and transfer learning, providing new perspectives on interdisciplinary. Experimental results demonstrate that CITL achieves competitive accuracies of 76.32% and 73.15% for detecting autism spectrum disorder and attention deficit hyperactivity disorder, respectively, which outperforms existing related transfer learning work for 7.2% and 0.5% respectively.

Paper Structure

This paper contains 13 sections, 13 equations, 5 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Methodology
  3. dFCs construction
  4. TransferNet
  5. Pseudo Label Generative Model and Enhanced Representation Generator
  6. Classification
  7. Implementation details
  8. Results
  9. Datasets and Preprocessing
  10. Competing methods
  11. Classification Performance
  12. Ablation studies
  13. Conclusion

Figures (5)

  • Figure 1: illustrates the proposed CITL framework, which consists of three main components: (a) Construction of dFCs, (b) Pre-trained TransferNet, which is used to screen out the dFCs that possess comorbid patterns, (c) Enhanced Representation Generator and Classification which contains an encoder-decoder generator and an architecturally simple classifier.
  • Figure 2: illustrates the two convolution modules of the Depthwise Separable Convolution Model: (a) represents the ColumnFeatureConv Model, which primarily focuses on edge information, and (b) shows the RowFeatureConv module, which is used to integrate node features.
  • Figure 3: The Conversion Engine removes the set with fewer samples and then optimizes the remaining dFCs through average pooling. The dFCs are transformed into the Reconstruction FC
  • Figure 4: The Feature Optimization AE model. The variables from the hidden layer are transformed into a 6670-dimensional vector, which represents the upper triangular part of the 116 templates and is used to construct the upper triangular part of the matrix. This matrix is transposed, and its diagonal elements are set to 1, resulting in a symmetric matrix referred to as the Optimization FC.
  • Figure 5: The ablation experiments of different modules. The result is shown in mean±std over 10-fold cross validation. On the x-axis, '1' represents the results of ablation experiment (1), '2' represents the results of ablation experiment (2) and '3' represents the results of ablation experiment (3).