Reference-Guided Pseudo-Label Generation for Medical Semantic Segmentation
Authors: Constantin Marc Seibold, Simon Reiß, Jens Kleesiek, Rainer Stiefelhagen2171-2179
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on two medical tasks, namely chest radiograph anatomy segmentation and retinal fluid segmentation. We achieve the same performance as a standard fully supervised model on X-ray anatomy segmentation, albeit 95% fewer labeled images. Aside from an in-depth analysis of different aspects of our proposed method, we further demonstrate the effectiveness of our reference-guided learning paradigm by comparing our approach against existing methods for retinal fluid segmentation with competitive performance as we improve upon recent work by up to 15% mean Io U. We demonstrate the effectiveness of our methods with extensive experiments for multi-class and binary multi-class semi-supervised semantic segmentation on the RETOUCH (Bogunovi c et al. 2019) and JSRT (Shiraishi et al. 2000) datasets. We provide a detailed ablation study investigating different aspects of our pseudo-labels in various settings. |
| Researcher Affiliation | Academia | Constantin Marc Seibold,1 Simon Reiß,1 Jens Kleesiek,2 Rainer Stiefelhagen1 1Institute for Anthropomatics and Robotics, Karlsruhe Institute of Technology, Germany 2Institute for Artificial Intelligence in Medicine, University Medicine Essen, Germany constantin.seibold@kit.edu, simon.reiss@kit.edu, jens.kleesiek@uk-essen.de, rainer.stiefelhagen@kit.edu |
| Pseudocode | No | The paper does not contain a clearly labeled pseudocode or algorithm block. |
| Open Source Code | No | The paper does not provide an explicit statement about releasing source code or a link to a code repository. |
| Open Datasets | Yes | For multi-label anatomy segmentation, we employ the public JSRT-dataset (Shiraishi et al. 2000). For multi-class retinal fluid segmentation, we utilize the Spectralis vendor of the RETOUCH data set consisting of 14 optical coherence tomography volumes with 49 b-scans each. |
| Dataset Splits | Yes | For each amount of labeled images, we choose to generate five distinct random splits from the first set using Nl labeled images (Nl 3, 6, 12, 24). For each split, we use five images of the first set for validation while using the second set for testing. We follow the setup of (Reiß et al. 2021) and thus perform 10-fold cross-validation with training sets using Nl labeled images (Nl 3, 6, 12, 24), with validation and test sets of roughly equal size on Spectralis |
| Hardware Specification | Yes | All experiments were run on one 11GB NVIDIA Ge Force RTX 2080. |
| Software Dependencies | No | The paper does not specify software dependencies with version numbers (e.g., Python version, specific library versions). |
| Experiment Setup | Yes | We optimize using Adam (Kingma and Ba 2014) with learning rate and weight decay of 0.0005 for 100 epochs on JSRT, 200 on extended JSRT and 50 epochs on RETOUCH respectively. As data augmentations, we use random cropping, rotation, additive noise, and color jitters with additional random flipping. For JSRT, we use batch size 5 with image size 512, while for RETOUCH we use batch size 8 following the preprocessing utilized in (Reiß et al. 2021). For all experiments, we build each batch as a combination of P with p = 3 and randomly sampled images of the whole dataset. We set the number of considered nearest neighbors k = 7000 and the representation map size s = 64. |