MSSDA: Multi-Sub-Source Adaptation for Diabetic Foot Neuropathy Recognition

TL;DR

This work addresses DFN recognition from plantar pressure by introducing a continuous plantar-pressure dataset (DFN-DS) and a novel three-stage multi-source domain adaptation method (MSSDA). MSSDA partitions source data into pseudo-sub-domains via convolutional feature-statistics, selects a subset of sub-domains closest to the target, and aligns each pair of source-target domains across multiple feature spaces to minimize domain shift while avoiding negative transfer. The approach achieves superior accuracy and recall on DFN-DS and the FRA dataset, outperforming a broad set of baselines and demonstrating robust cross-subject and cross-dataset performance. The work provides a practical resource and a scalable framework for cross-subject DFN recognition with potential for real-world screening and monitoring.

Abstract

Diabetic foot neuropathy (DFN) is a critical factor leading to diabetic foot ulcers, which is one of the most common and severe complications of diabetes mellitus (DM) and is associated with high risks of amputation and mortality. Despite its significance, existing datasets do not directly derive from plantar data and lack continuous, long-term foot-specific information. To advance DFN research, we have collected a novel dataset comprising continuous plantar pressure data to recognize diabetic foot neuropathy. This dataset includes data from 94 DM patients with DFN and 41 DM patients without DFN. Moreover, traditional methods divide datasets by individuals, potentially leading to significant domain discrepancies in some feature spaces due to the absence of mid-domain data. In this paper, we propose an effective domain adaptation method to address this proplem. We split the dataset based on convolutional feature statistics and select appropriate sub-source domains to enhance efficiency and avoid negative transfer. We then align the distributions of each source and target domain pair in specific feature spaces to minimize the domain gap. Comprehensive results validate the effectiveness of our method on both the newly proposed dataset for DFN recognition and an existing dataset.

Figures (3)

• Figure 1: Unlike conventional methods, our method aims to discover latent domains and use the source data with some specific domain labels to train the model but not all of them.
• Figure 2: An overview of the proposed three-stage framework MSSDA. $F_i$ denotes a feature extractor, $D_i$ signifies a domain discriminator, $C_i$ represents a classifier. Note that the $F_0$ trained in stage 1 is used in stage 2 while $F_1$,..., $F_M$ are not fine-tuned from $F_0$. In stage 3, there are specific feature extractors, domain discriminator and classifiers for each selected source domain. Please be aware that $F_0$, $F_1$,..., $F_M$ neither share the network architecture nor the weights. (Best viewed in color.)
• Figure 3: Results of the accuracy using other thresholds on DFN-DS.