Quick Search:

SE-BioNet: A Squeeze-and-Excitation CNN for Non-Invasive Overthinking Detection Using Biofield Imaging

Gornale, Shivanand, Patil, Supriya, Siddalingappa, Rashmi ORCID logoORCID: https://orcid.org/0000-0001-9786-8436, Deo, Anup W. and Athani, Abdulgaphur (2026) SE-BioNet: A Squeeze-and-Excitation CNN for Non-Invasive Overthinking Detection Using Biofield Imaging. F1000 Research.

[thumbnail of 4bfbeba0-c179-4734-9115-1e9a39607bc0_f1000res_192292.pdf]
Preview
Text
4bfbeba0-c179-4734-9115-1e9a39607bc0_f1000res_192292.pdf - Published Version
Available under License Creative Commons Attribution.

| Preview

Abstract

Background
The human biofield has been explored as a possible indicator of physical and emotional states. Complementary and alternative medicine (CAM) practices, such as Reiki, Therapeutic Touch, and Pranic Healing, incorporate biofield-related information into their assessment and therapeutic approaches. Gas Discharge Visualization (GDV) and Poly Contrast Interference Photography (PIP) are among the imaging techniques used to visualize biofield patterns. However, most existing studies have focused on GDV images, with comparatively limited research using PIP images and deep learning techniques. Overthinking is a growing mental health concern that is often difficult to identify using conventional diagnostic methods. This study explores whether deep learning can identify patterns in PIP biofield images associated with overthinking.

Methods
A convolutional neural network (CNN) integrated with a Squeeze-and-Excitation (SE) attention mechanism was developed for binary classification of biofield images into two classes: overthinking and no overthinking pattern. The SE module enhances discriminative feature learning by adaptively recalibrating channel-wise feature responses. The model was trained and evaluated separately on datasets annotated independently by two experts (Expert-I and Expert-II), and its performance was assessed using 5-fold cross-validation.

Results
The proposed SE-BioNet model achieved classification accuracies of 93.95% and 93.80% on Expert-I and Expert-II annotated datasets, respectively, with corresponding F1-scores of 94.02% and 93.59%. Five-fold cross-validation yielded mean accuracies of 92.22% ± 1.17% and 92.08% ± 1.34%, demonstrating stable generalisation.

Conclusions
The proposed framework demonstrates technical feasibility and consistent performance across cross-validation folds. However, validation using larger, multi-centre datasets is necessary to assess its generalizability. These findings highlight the potential of attention-based model in learning more useful features though further testing with larger datasets and expert validation is necessary for clinical application.

Item Type: Article
Status: Published
DOI: 10.12688/f1000research.192292.1
School/Department: London Campus
URI: https://ray.yorksj.ac.uk/id/eprint/16241

University Staff: Request a correction | RaY Editors: Update this record