A New Noisy Label Learning Method For Transdiagnostic Classification of Mental Disorders
Poster No:
1469
Submission Type:
Abstract Submission
Authors:
Yuhui Du1, Zheng Wang1, Vince Calhoun2
Institutions:
1Shanxi University, Taiyuan, Shanxi, 2GSU/GATech/Emory, Atlanta, GA
First Author:
Co-Author(s):
Introduction:
The subjective nature of diagnosing mental disorders complicates achieving accurate diagnoses(David & Deeley, 2024). The overlap of symptoms among various disorders exacerbates this issue, particularly in clinical practice where conditions like bipolar disorder (BP) and schizophrenia (SZ) can present similar clinical symptoms and cognitive impairments. This increases the risk of misdiagnosing labels. Research on learning algorithms for noisy labels in the field of image classification can help address this issue in the diagnosis of mental illnesses(Cordeiro, Sachdeva, Belagiannis, Reid, & Carneiro, 2023; Patel & Sastry, 2023), hence we propose a new label learning algorithm.
Methods:
We propose a new deep learning-based method named mutualistic multi-network noisy label learning (MMNNLL) for robust learning from data with noisy labels, aimed at mitigating potential diagnostic bias in mental disorders. This approach moves beyond clinical disorder classifications and contributes to the search for precise biomarkers from a transdiagnostic perspective. We develop a multi-network architecture that integrates independently trained deep neural networks (DNNs) and their mean-teacher DNNs to ensure prediction consistency across samples with both clean and noisy labels. Our method iteratively identifies samples with clean labels while also leveraging information from data with noisy labels, thus enabling robust model construction despite the presence of potential label noise.
On the CIFAR-10 dataset, our method was compared with several noise-label learning algorithms. Given the potential shared characteristics between BP and SZ, we analyzed resting-state fMRI data of 184 healthy controls (HCs), 107 BP patients with psychosis, and 170 SZ patients from the Bipolar and Schizophrenia Network on Intermediate Phenotypes (BSNIP) project to explore transdiagnostic classes (also known as neuroimaging-driven biotypes) among BP and SZ patients using our MMNNLL method, and then investigated their group differences as well as unique brain functional changes relative to HCs.
On the CIFAR-10 dataset, our method was compared with several noise-label learning algorithms. Given the potential shared characteristics between BP and SZ, we analyzed resting-state fMRI data of 184 healthy controls (HCs), 107 BP patients with psychosis, and 170 SZ patients from the Bipolar and Schizophrenia Network on Intermediate Phenotypes (BSNIP) project to explore transdiagnostic classes (also known as neuroimaging-driven biotypes) among BP and SZ patients using our MMNNLL method, and then investigated their group differences as well as unique brain functional changes relative to HCs.
Results:
Our method achieves high accuracy under various noise conditions in the public image dataset CIFAR-10. Under symmetric label noise, our method achieved a classification accuracy of 90.4% with a 20% label noise rate and 88.35% with a 40% label noise rate on the CIFAR-10 dataset. Furthermore, under pair flipping noise at a 40% label noise rate, our method reached an accuracy of 79.15%. In summary, models trained with our approach show significant improvement in classification performance on the test set compared to other methods across different types and levels of label noise.
The experiments on the BSNIP dataset also showed very good results. Fig. 1 reveals that the T-value differences in Functional Network Connectivity (FNC) between two biotypes are more significant than those between traditional diagnostic categories, with more pronounced brain changes in these biotypes compared to Bipolar Disorder (BP) and Schizophrenia (SZ), highlighting their biological importance. Biotype 1 and Biotype 2 show substantial FNC differences, particularly in sensorimotor, subcortical, visual, cognitive control, default mode, and cerebellum domains. Fig. 2 compares mean connectivity strengths for FNCs in BP, SZ, and Healthy Controls (HCs), and within biotypes, showing more pronounced differences between Biotype 1 and Biotype 2 than between BP and SZ. Despite BP-SZ differences not surviving correction, 47 FNCs show significant differences between biotypes post-Bonferroni correction, compared to 18 for BP vs. SZ (p<0.005), indicating the robustness of biotype distinctions.
The experiments on the BSNIP dataset also showed very good results. Fig. 1 reveals that the T-value differences in Functional Network Connectivity (FNC) between two biotypes are more significant than those between traditional diagnostic categories, with more pronounced brain changes in these biotypes compared to Bipolar Disorder (BP) and Schizophrenia (SZ), highlighting their biological importance. Biotype 1 and Biotype 2 show substantial FNC differences, particularly in sensorimotor, subcortical, visual, cognitive control, default mode, and cerebellum domains. Fig. 2 compares mean connectivity strengths for FNCs in BP, SZ, and Healthy Controls (HCs), and within biotypes, showing more pronounced differences between Biotype 1 and Biotype 2 than between BP and SZ. Despite BP-SZ differences not surviving correction, 47 FNCs show significant differences between biotypes post-Bonferroni correction, compared to 18 for BP vs. SZ (p<0.005), indicating the robustness of biotype distinctions.
Conclusions:
In conclusion, to address the potential biases or errors in clinical diagnosis of mental illnesses, we propose a mutualistic multi-network noisy label learning method that optimizes the consistency of multiple DNNs in identifying and utilizing samples with both clean and noisy labels, thus learning from data with noisy labels.
Disorders of the Nervous System:
Psychiatric (eg. Depression, Anxiety, Schizophrenia) 2
Modeling and Analysis Methods:
fMRI Connectivity and Network Modeling 1
Methods Development
Keywords:
FUNCTIONAL MRI
Other - Noisy Label Learning, Multi-network, Biotypes, Bipolar Disorder, Schizophrenia
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
Provide references using APA citation style.
Pearlson, G. D., Clementz, B. A., Sweeney, J. A., Keshavan, M. S., & Tamminga, C. A. J. P. C. (2016). Does biology transcend the symptom-based boundaries of psychosis? , 39(2), 165-174.
Yamada, Y., Matsumoto, M., Iijima, K., & Sumiyoshi, T. J. C. p. d. (2020). Specificity and continuity of schizophrenia and bipolar disorder: relation to biomarkers. 26(2), 191-200.