BioSpectrum: A Human-in-the-Loop Framework Combining Convolutional Neural Networks and Conversational AI for Accessible Medical Image Triage

DOI

10.2991/978-94-6239-821-4_9How to use a DOI?

Keywords

Convolutional neural networks; medical image classification; vision-language models; zero-shot learning; clinical decision support; human-in-the-loop AI; automation bias; explainable AI; Grad-CAM; healthcare access; dermatology AI; telemedicine; digital health

Abstract

Access to specialist medical diagnosis remains highly unequal worldwide: the World Health Organization projects a global shortfall of roughly 11 million health workers by 2030, concentrated overwhelmingly in low- and middle-income regions [1]. Convolutional neural networks (CNNs) have repeatedly matched specialist-level accuracy on narrow, single-modality image classification tasks such as dermoscopic skin lesion classification [2], but these systems remain confined largely to research settings due to dataset, infrastructure, and validation barriers. Separately, general-purpose vision-language models (VLMs) such as GPT-4V and Claude have demonstrated an ability to reason over medical images without task-specific training, but recent evaluations show this zero-shot capability is inconsistent and, on tasks such as radiological pathology identification, has been measured at accuracies as low as 35% [3], well below what would be acceptable for autonomous diagnosis. This paper proposes BioSpectrum, a system architecture that deliberately separates these two capabilities rather than conflating them: a modality-specific, transfer-learned CNN performs the actual classification task and is evaluated against ground-truth clinical labels, while a VLM-based conversational layer explains the CNN's output, retrieves relevant recent literature, and answers follow-up questions — without altering the underlying classification. A mandatory clinician confirmation gate sits between model output and any recorded diagnosis, directly addressing the documented risk of automation bias in AI-assisted clinical decision-making [4]. We describe a two-tier deployment model: a public referral tool that routes flagged users to verified doctors and clinics without issuing a diagnosis, and a hospital-facing dashboard that supports the full clinical workflow including prescription drafting, pharmacy stock lookup, billing, and longitudinal patient records. We close with an explicit account of the system's current limitations — including dataset bias against darker skin tones documented in existing dermatology AI literature [5,6], the gap between benchmark and real-world accuracy, and the regulatory pathway required before any clinical deployment — framing these as the defined agenda for future validation work rather than as solved problems.

Copyright

© 2026 The Author(s)

Open Access

Open Access This chapter is licensed under the terms of the Creative Commons Attribution-NonCommercial 4.0 International License (http://creativecommons.org/licenses/by-nc/4.0/), which permits any noncommercial use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.

Download article (PDF)

TY  - CONF
AU  - Suyash Meena
AU  - Preenon Bagchi
PY  - 2026
DA  - 2026/10/31
TI  - BioSpectrum: A Human-in-the-Loop Framework Combining Convolutional Neural Networks and Conversational AI for Accessible Medical Image Triage
BT  - Proceedings of the International Conference on Information Systems Design And Intelligent Applications (ICISDA 2026)
PB  - Atlantis Press
SP  - 89
EP  - 97
SN  - 1951-6851
UR  - https://doi.org/10.2991/978-94-6239-821-4_9
DO  - 10.2991/978-94-6239-821-4_9
ID  - Meena2026
ER  -