JMIR Publications

ABSTRACT

Background:

Neglected tropical diseases (NTDs) are the most prevalent diseases comprising of 21 different conditions, where more than half of these conditions have skin manifestations, skin NTDs. However, little has been done towards the utilization of deep learning (DL)-based diagnostic models based on multimodal data fusion (MMDF) techniques for skin NTDs.

Objective:

This article, thus, presents a systematic review of the DL-based MMDF methods for the diagnosis of skin NTDs.

Methods:

The PRISMA method is used to conduct the review on both skin NTDs and non-NTD skin diseases due to insufficient studies demonstrating MMDF for skin NTDs. Additionally, the ethical and potential risk of biases were analyzed.

Results:

Accordingly, 437 articles were initially collected from seven major and reputed sources, where 14 articles are finally selected and critically appraised using six parameters (research approaches followed, disease(s) diagnosed, dataset(s) used, algorithm(s) applied, and performance achieved). Results suggest that the MMDF methods have improved diagnostic model performances of both skin NTDs and non-NTD skin diseases where, algorithmically, CNN-based models are the predominantly used DL architectures (71.5% studies) serving disease classification, image feature extraction when used with methods like transformer-based methods (14.25% studies) including generative models (14.25% studies).

Conclusions:

This article suggests further studies on model efficiency, data scarcity, algorithm selection and utilization, and fusion strategies of multiple modalities, because of the major gaps of the limited investigations of DL-based MMDF for skin NTDs, which hinders their potential adoption in resource-constrained areas.