Biologically-inspired Semi-supervised Semantic Segmentation for Biomedical Imaging

TL;DR

The paper tackles semantic segmentation in biomedical imaging under data scarcity by proposing a two-stage pipeline that first uses biologically inspired Hebbian learning to pretrain both convolutional and transpose-convolutional layers in an unsupervised manner, then fine-tunes the model with backpropagation on a small labeled set. It introduces novel Hebbian rules for transpose-convolution layers and provides two strategies (SWTA-S/HPCA-S and SWTA-TSA/HPCA-TSA) to handle the upsampling path. Extensive experiments on GlaS, PH2, HMEPS, and LA XIONG demonstrate state-of-the-art or near-state-of-the-art performance across varying label availability, and show that the unsupervised stage can further boost existing semi-supervised methods when used for initialization. The approach offers a practical, energy-efficient, and biologically plausible alternative for biomedical image segmentation where annotated data is expensive to obtain.

Abstract

We propose a novel bio-inspired semi-supervised learning approach for training downsampling-upsampling semantic segmentation architectures. The first stage does not use backpropagation. Rather, it exploits the Hebbian principle ``fire together, wire together'' as a local learning rule for updating the weights of both convolutional and transpose-convolutional layers, allowing unsupervised discovery of data features. In the second stage, the model is fine-tuned with standard backpropagation on a small subset of labeled data. We evaluate our methodology through experiments conducted on several widely used biomedical datasets, deeming that this domain is paramount in computer vision and is notably impacted by data scarcity. Results show that our proposed method outperforms SOTA approaches across different levels of label availability. Furthermore, we show that using our unsupervised stage to initialize the SOTA approaches leads to performance improvements. The code to replicate our experiments can be found at https://github.com/ciampluca/hebbian-bootstraping-semi-supervised-medical-imaging

Figures (7)

• Figure 1: Our semi-supervised, bio-inspired approach for semantic segmentation architectures with downsampling and upsampling paths. The method unfolds in two stages. First, we employ Hebbian learning for unsupervised pre-training on a large set of unlabeled data, enabling the model to autonomously identify features like cluster centroids. Notably, we derive new Hebbian learning rules for the transpose-convolution layers in the upsampling path. In the second stage, we apply supervised backpropagation to fine-tune the model using a small labeled dataset.
• Figure 2: Building blocks for Hebbian learning. Hebbian updates are computed from the difference between a target signal $\mathbf{s}$ and a reconstructed signal $\mathbf{s*}$, and a gating signal $\mathbf{g}$ which modulates the update steps $\mathbf{g}$. Gating and reconstruction signals are, in turn, derived from outputs and weights through the respective blocks.
• Figure 3: Hebbian learning in convolutional and transpose-convolutional layers.
• Figure 4: Performance changes (with green and red arrows) obtained with single-stage SOTA semi-supervised approaches initialized with our unsupervised Hebbian pre-training compared to initialization from scratch (blue bar). Each row corresponds to a dataset, while each column to a different degree of label availability. We report DC values embedded in the most convenient range for best readability.
• Figure 5: Qualitative results from our semi-supervised approach based on Hebbian SWTA-TSA. Each row corresponds to a different percentage of label availability; each column corresponds to a different dataset and includes a triplet sample-target-prediction.