AFiRe: Anatomy-Driven Self-Supervised Learning for Fine-Grained Representation in Radiographic Images

Authors: Yihang Liu, Lianghua He, Ying Wen, Longzhen Yang, Hongzhou Chen

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that AFi Re: (i) provides robust anatomical discrimination, achieving more cohesive feature clusters compared to state-of-the-art contrastive learning methods; (ii) demonstrates superior generalization, surpassing 7 radiography-specific self-supervised methods in multi-label classification tasks with limited labeling; and (iii) integrates fine-grained information, enabling precise anomaly detection using only image-level annotations. Our extensive experiments demonstrate that AFi Re enhances the model s capacity to capture fine-grained discriminative information from normal radiographic images, facilitating a more robust representation across different anatomic structures. Compared to different supervised and self-supervised benchmarks, AFi Re achieves superior performance in multi-label classification tasks, indicating its generalization in analyzing real disease images.
Researcher Affiliation Academia 1 School of Computer Science and Technology, Tongji University, Shanghai, China. 2 Shanghai Eye Diseases Prevention and Treatment Center, Shanghai Eye Hospital, Shanghai, China. 3 School of Communication and Electronic Engineering, East China Normal University, Shanghai, China. EMAIL, EMAIL
Pseudocode No The paper describes the methodology using textual explanations and figures, but no explicitly labeled pseudocode or algorithm blocks are present.
Open Source Code Yes Code https://github.com/LYH-hh/AFi Re
Open Datasets Yes In this paper, we assemble a dataset of 811,170 Chest X-ray (CXR) images for pre-training, which includes 81,117 normal CXR images sourced from MIMIC-CXR-JPG (Johnson et al. 2019) and 81, 117 9 synthetic abnormal images. We evaluate our model on three CXR datasets: (i) Child CXRs (Kermany et al. 2018), (ii) NIH (Wang et al. 2017), (iii) CXP (Irvin et al. 2019), and (iv) SIIM-ARC 1, using the official data partitions.
Dataset Splits Yes We evaluate our model on three CXR datasets: (i) Child CXRs (Kermany et al. 2018), (ii) NIH (Wang et al. 2017), (iii) CXP (Irvin et al. 2019), and (iv) SIIM-ARC 1, using the official data partitions. To ensure robustness, we further conduct five-fold cross-validation for these experiments. AUC (%) for multi-label classification on NIH and CXP, and Dice (%) for segmentation on SIIM-ACR are reported. The best results are bolded. 5-C denotes the five-fold cross-validation.
Hardware Specification Yes We implement our pre-training model using Py Torch and distribute the training across 8 NVIDIA A6000 GPUs, running for 800 epochs with a batch size of 64 for 6 days.
Software Dependencies No The paper states 'We implement our pre-training model using Py Torch' but does not specify a version number for PyTorch or any other software dependencies.
Experiment Setup Yes We follow the DINO (Caron et al. 2021) training paradigm (e.g., optimizer, learning rate, and weight decay schedule), utilizing an input image size of 224 224. ...running for 800 epochs with a batch size of 64 for 6 days. ... τ is the temperature parameter and H(q, c) = PK k=1 q(k) log c(k), which is the cross-entropy function to measure the similarity between two probabilities. ...through the momentum parameter m.