Zilin Lu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Zilin Lu has a total of 53 co-authors. Here are the top co-authors by number of shared publications:
- Yong Xia
School of Mathematical Sciences Beihang University 100191 Beijing People's Republic of China
10 shared publications - Qingjie Zeng
National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, School of Computer Science and Engineering, Northwestern Polytechnical University, Xian, China
9 shared publications - Yutong Xie
Mohamed bin Zayed University of Artificial Intelligence
7 shared publications - Mengkang Lu
National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, School of Computer Science and Engineering, Northwestern Polytechnical University, Xian, China
5 shared publications Huan Luo
National Engineering Laboratory for Integrated Aero-Space-Ground-Ocean Big Data Application Technology, School of Computer Science and Engineering, Northwestern Polytechnical University, Xian, China
2 shared publications- Yanning Zhang
Northwestern Polytechnical University
2 shared publications
Latest Publications
- 17 September 2026
- Published by American Association for the Advancement of Science (AAAS) in Science
- Vol. 393 (6817)
An expert-level generalist AI for abdominal CT diagnosis
Artificial intelligence (AI) in radiology aspires to deliver expert-level diagnosis across diverse clinical tasks, yet existing supervised strategies remain limited in scope. We developed RADAR, a generalist vision-language model trained on more than 400,000 contrast-enhanced abdominal computed tomography (CT) examinations and 15 million anatomy-wise image-text pairs, learning directly from clinical reports without manual annotation. Throughout internal and external evaluations across multiple centers and varied clinical scenarios, RADAR achieved high diagnostic performance and robust generalization for 18 anatomical structures and 146 imaging findings. In a reader study, RADAR assistance increased the diagnostic sensitivity of 26 radiologists by ~10%. RADAR offers a scalable, versatile, and interpretable solution for abdominal CT, demonstrating that generalist AI can match human experts in general and complicated radiology tasks.
- 17 September 2026
- Published by American Association for the Advancement of Science (AAAS) in Science
- Vol. 393 (6817)
An expert-level generalist AI for abdominal CT diagnosis
Artificial intelligence (AI) in radiology aspires to deliver expert-level diagnosis across diverse clinical tasks, yet existing supervised strategies remain limited in scope. We developed RADAR, a generalist vision-language model trained on more than 400,000 contrast-enhanced abdominal computed tomography (CT) examinations and 15 million anatomy-wise image-text pairs, learning directly from clinical reports without manual annotation. Throughout internal and external evaluations across multiple centers and varied clinical scenarios, RADAR achieved high diagnostic performance and robust generalization for 18 anatomical structures and 146 imaging findings. In a reader study, RADAR assistance increased the diagnostic sensitivity of 26 radiologists by ~10%. RADAR offers a scalable, versatile, and interpretable solution for abdominal CT, demonstrating that generalist AI can match human experts in general and complicated radiology tasks.
- Top Cited Papers
- 20 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 35 (10577149) , 3228-3239
Decoupling Target Semantics via Text-Anchored Visual Contrast for Semi-Supervised Medical Image Segmentation
Semi-supervised learning (SSL) provides an effective means of reducing reliance on large-scale annotated datasets by leveraging unlabeled data. However, existing SSL methods often struggle with semantic ambiguity, especially under limited supervision. Recent studies have incorporated textual information to provide contextual guidance, yet most focus on feature fusion rather than emphasizing target semantics critical for segmentation. In this paper, we proposed a novel Text-anchored Visual Decoupling (TeViD) framework for semi-supervised medical image segmentation. TeViD is built upon a teacher-student architecture with a dual-decoder design that explicitly disentangles target and background representations using both labeled and unlabeled data. For unlabeled data, a reversed cross-supervision mechanism is introduced to enhance decoder diversity and semantic separation. Furthermore, two contrastive learning objectives are proposed: a teacher-guided visual contrastive loss and a text-anchored contrastive loss, both designed to reinforce semantic disentanglement from visual and textual perspectives. Extensive experiments on five public datasets (covering X-ray, pathology, ultrasound, MRI, and CT) demonstrate that TeViD consistently outperforms both standard SSL and text-enhanced SSL methods, achieving average improvements of 5.72% in Dice and 8.15% in mIoU over the second-best competitor. The code is available at: https://github.com/jgfiuuuu/TeViD.
- 20 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 35 (10577149) , 3228-3239
Top Cited PapersDecoupling Target Semantics via Text-Anchored Visual Contrast for Semi-Supervised Medical Image Segmentation
Semi-supervised learning (SSL) provides an effective means of reducing reliance on large-scale annotated datasets by leveraging unlabeled data. However, existing SSL methods often struggle with semantic ambiguity, especially under limited supervision. Recent studies have incorporated textual information to provide contextual guidance, yet most focus on feature fusion rather than emphasizing target semantics critical for segmentation. In this paper, we proposed a novel Text-anchored Visual Decoupling (TeViD) framework for semi-supervised medical image segmentation. TeViD is built upon a teacher-student architecture with a dual-decoder design that explicitly disentangles target and background representations using both labeled and unlabeled data. For unlabeled data, a reversed cross-supervision mechanism is introduced to enhance decoder diversity and semantic separation. Furthermore, two contrastive learning objectives are proposed: a teacher-guided visual contrastive loss and a text-anchored contrastive loss, both designed to reinforce semantic disentanglement from visual and textual perspectives. Extensive experiments on five public datasets (covering X-ray, pathology, ultrasound, MRI, and CT) demonstrate that TeViD consistently outperforms both standard SSL and text-enhanced SSL methods, achieving average improvements of 5.72% in Dice and 8.15% in mIoU over the second-best competitor. The code is available at: https://github.com/jgfiuuuu/TeViD.
- 1 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 29 (12) , 8976-8987
PathBot: A Foundation Model for Pathological Image Analysis
Computational pathology has emerged as a transformative paradigm by leveraging artificial intelligence to automate and enhance diagnostic procedures. However, existing models often target narrow tasks or specific tumor types, missing opportunities to unify diverse datasets and tasks through joint learning. In this work, we introduce PathBot, a foundation model tailored for comprehensive pathological image analysis. Central to PathBot is a ViTGiant encoder with one billion parameters, the largest model to date trained on publicly available pathological data. We pre-train this encoder using a novel Masked Distillation Network (MDN) and an integrated learning strategy that combines contrastive and generative objectives. The pretraining leverages over 30 million image patches derived from 11,765 whole slide images (WSIs) across 32 cancer types in the Cancer Genome Atlas (TCGA). To evaluate its versatility, we pair the encoder with task-specific decoders for segmentation, detection, classification, and regression. Extensive experiments across 20 downstream tasks demonstrate that PathBot achieves state-of-the-art performance in most cases, showcasing its robustness and generalizability. Code and models will be released to support further research.
- 1 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 29 (12) , 8976-8987
PathBot: A Foundation Model for Pathological Image Analysis
Computational pathology has emerged as a transformative paradigm by leveraging artificial intelligence to automate and enhance diagnostic procedures. However, existing models often target narrow tasks or specific tumor types, missing opportunities to unify diverse datasets and tasks through joint learning. In this work, we introduce PathBot, a foundation model tailored for comprehensive pathological image analysis. Central to PathBot is a ViTGiant encoder with one billion parameters, the largest model to date trained on publicly available pathological data. We pre-train this encoder using a novel Masked Distillation Network (MDN) and an integrated learning strategy that combines contrastive and generative objectives. The pretraining leverages over 30 million image patches derived from 11,765 whole slide images (WSIs) across 32 cancer types in the Cancer Genome Atlas (TCGA). To evaluate its versatility, we pair the encoder with task-specific decoders for segmentation, detection, classification, and regression. Extensive experiments across 20 downstream tasks demonstrate that PathBot achieves state-of-the-art performance in most cases, showcasing its robustness and generalizability. Code and models will be released to support further research.
- 21 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 45 (2) , 477-489
Harnessing Text Insights With Visual Alignment for Medical Image Segmentation
Pre-trained vision-language models (VLMs) and language models (LMs) have recently garnered significant attention due to their remarkable ability to represent textual concepts, opening up new avenues in vision tasks. In medical image segmentation, efforts are being made to integrate text and image data using VLMs and LMs. However, current text-enhanced approaches face several challenges. First, using separate pre-trained vision and text models to encode image and text data can result in semantic shifts. Second, while VLMs can establish the correspondence between visual and textual features when pre-trained on paired image-text data, this alignment often deteriorates during segmentation tasks due to misalignment between the text and vision components in ongoing learning. In this paper, we propose TeViA, a novel approach that seamlessly integrates with various vision and text models, irrespective of their pre-training relationships. This integration is achieved through a segmentation-specific text-to-vision alignment design, ensuring both information gain and semantic consistency. Specifically, for each training data, a foreground visual representation is extracted from the segmentation head and used to supervise projection layers, thereby adjusting the textual features to better contribute to the segmentation task. Additionally, a historic visual prototype is created by aggregating target semantics from all training data and is updated using a momentum-based manner. This prototype aims to enhance the visual representation of each data instance by establishing feature-level connections, which in turn refines the textual features. The superiority of TeViA is validated on five public datasets, exhibiting over 6% Dice improvements compared to vision-only methods. Code is available at: https://github.com/jgfiuuuu/TeViA.
- 21 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 45 (2) , 477-489
Harnessing Text Insights With Visual Alignment for Medical Image Segmentation
Pre-trained vision-language models (VLMs) and language models (LMs) have recently garnered significant attention due to their remarkable ability to represent textual concepts, opening up new avenues in vision tasks. In medical image segmentation, efforts are being made to integrate text and image data using VLMs and LMs. However, current text-enhanced approaches face several challenges. First, using separate pre-trained vision and text models to encode image and text data can result in semantic shifts. Second, while VLMs can establish the correspondence between visual and textual features when pre-trained on paired image-text data, this alignment often deteriorates during segmentation tasks due to misalignment between the text and vision components in ongoing learning. In this paper, we propose TeViA, a novel approach that seamlessly integrates with various vision and text models, irrespective of their pre-training relationships. This integration is achieved through a segmentation-specific text-to-vision alignment design, ensuring both information gain and semantic consistency. Specifically, for each training data, a foreground visual representation is extracted from the segmentation head and used to supervise projection layers, thereby adjusting the textual features to better contribute to the segmentation task. Additionally, a historic visual prototype is created by aggregating target semantics from all training data and is updated using a momentum-based manner. This prototype aims to enhance the visual representation of each data instance by establishing feature-level connections, which in turn refines the textual features. The superiority of TeViA is validated on five public datasets, exhibiting over 6% Dice improvements compared to vision-only methods. Code is available at: https://github.com/jgfiuuuu/TeViA.
- 17 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 44 (11) , 4605-4616
Bridging the Semantic Gap in Medical Visual Question Answering With Prompt Learning
Medical Visual Question Answering (Med-VQA) aims to answer questions regarding the content of medical images, crucial for enhancing diagnostics and education in healthcare. However, progress in this field is hindered by data scarcity due to the resource-intensive nature of medical data annotation. While existing Med-VQA approaches often rely on pre-training to mitigate this issue, bridging the semantic gap between pre-trained models and specific tasks remains a significant challenge. This paper presents the Dynamic Semantic-Adaptive Prompting (DSAP) framework, leveraging prompt learning to enhance model performance in Med-VQA. To this end, we introduce two prompting strategies: Semantic Alignment Prompting (SAP) and Dynamic Question-Aware Prompting (DQAP). SAP prompts multi-modal inputs during fine-tuning, reducing the semantic gap by aligning model outputs with domain-specific contexts. Simultaneously, DQAP enhances answer selection by leveraging grammatical relationships between questions and answers, thereby improving accuracy and relevance. The DSAP framework was pre-trained on three datasets—ROCO, MedICaT, and MIMIC-CXR—and comprehensively evaluated against 15 existing Med-VQA models on three public datasets: VQA-RAD, SLAKE, and PathVQA. Our results demonstrate a substantial performance improvement, with DSAP achieving a 1.9% enhancement in average results across benchmarks. These findings underscore DSAP’s effectiveness in addressing critical challenges in Med-VQA and suggest promising avenues for future developments in medical AI.
- 17 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 44 (11) , 4605-4616
Bridging the Semantic Gap in Medical Visual Question Answering With Prompt Learning
Medical Visual Question Answering (Med-VQA) aims to answer questions regarding the content of medical images, crucial for enhancing diagnostics and education in healthcare. However, progress in this field is hindered by data scarcity due to the resource-intensive nature of medical data annotation. While existing Med-VQA approaches often rely on pre-training to mitigate this issue, bridging the semantic gap between pre-trained models and specific tasks remains a significant challenge. This paper presents the Dynamic Semantic-Adaptive Prompting (DSAP) framework, leveraging prompt learning to enhance model performance in Med-VQA. To this end, we introduce two prompting strategies: Semantic Alignment Prompting (SAP) and Dynamic Question-Aware Prompting (DQAP). SAP prompts multi-modal inputs during fine-tuning, reducing the semantic gap by aligning model outputs with domain-specific contexts. Simultaneously, DQAP enhances answer selection by leveraging grammatical relationships between questions and answers, thereby improving accuracy and relevance. The DSAP framework was pre-trained on three datasets—ROCO, MedICaT, and MIMIC-CXR—and comprehensively evaluated against 15 existing Med-VQA models on three public datasets: VQA-RAD, SLAKE, and PathVQA. Our results demonstrate a substantial performance improvement, with DSAP achieving a 1.9% enhancement in average results across benchmarks. These findings underscore DSAP’s effectiveness in addressing critical challenges in Med-VQA and suggest promising avenues for future developments in medical AI.
- Top Cited Papers
- 31 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 44 (7) , 2948-2959
Segment Together: A Versatile Paradigm for Semi-Supervised Medical Image Segmentation
The scarcity of annotations has become a significant obstacle in training powerful deep-learning models for medical image segmentation, limiting their clinical application. To overcome this, semi-supervised learning that leverages abundant unlabeled data is highly desirable to enhance model training. However, most existing works still focus on specific medical tasks and underestimate the potential of learning across diverse tasks and datasets. In this paper, we propose a Versatile Semi-supervised framework (VerSemi) to present a new perspective that integrates various SSL tasks into a unified model with an extensive label space, exploiting more unlabeled data for semi-supervised medical image segmentation. Specifically, we introduce a dynamic task-prompted design to segment various targets from different datasets. Next, this unified model is used to identify the foreground regions from all labeled data, capturing cross-dataset semantics. Particularly, we create a synthetic task with a CutMix strategy to augment foreground targets within the expanded label space. To effectively utilize unlabeled data, we introduce a consistency constraint that aligns aggregated predictions from various tasks with those from the synthetic task, further guiding the model to accurately segment foreground regions during training. We evaluated our VerSemi framework against seven established SSL methods on four public benchmarking datasets. Our results suggest that VerSemi consistently outperforms all competing methods, beating the second-best method with a 2.69% average Dice gain on four datasets and setting a new state of the art for semi-supervised medical image segmentation. Code is available at https://github.com/maxwell0027/VerSemi.
- 31 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 44 (7) , 2948-2959
Top Cited PapersSegment Together: A Versatile Paradigm for Semi-Supervised Medical Image Segmentation
The scarcity of annotations has become a significant obstacle in training powerful deep-learning models for medical image segmentation, limiting their clinical application. To overcome this, semi-supervised learning that leverages abundant unlabeled data is highly desirable to enhance model training. However, most existing works still focus on specific medical tasks and underestimate the potential of learning across diverse tasks and datasets. In this paper, we propose a Versatile Semi-supervised framework (VerSemi) to present a new perspective that integrates various SSL tasks into a unified model with an extensive label space, exploiting more unlabeled data for semi-supervised medical image segmentation. Specifically, we introduce a dynamic task-prompted design to segment various targets from different datasets. Next, this unified model is used to identify the foreground regions from all labeled data, capturing cross-dataset semantics. Particularly, we create a synthetic task with a CutMix strategy to augment foreground targets within the expanded label space. To effectively utilize unlabeled data, we introduce a consistency constraint that aligns aggregated predictions from various tasks with those from the synthetic task, further guiding the model to accurately segment foreground regions during training. We evaluated our VerSemi framework against seven established SSL methods on four public benchmarking datasets. Our results suggest that VerSemi consistently outperforms all competing methods, beating the second-best method with a 2.69% average Dice gain on four datasets and setting a new state of the art for semi-supervised medical image segmentation. Code is available at https://github.com/maxwell0027/VerSemi.
- Top Cited Papers
- 5 January 2025
- Published by Springer Nature in International Journal of Computer Vision
- Vol. 133 (6) , 3296-3311
PICK: Predict and Mask for Semi-supervised Medical Image Segmentation
- 5 January 2025
- Published by Springer Nature in International Journal of Computer Vision
- Vol. 133 (6) , 3296-3311
Top Cited PapersPICK: Predict and Mask for Semi-supervised Medical Image Segmentation
- Top Cited Papers
- 1 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 44 (1) , 44-56
Consistency-Guided Differential Decoding for Enhancing Semi-Supervised Medical Image Segmentation
Semi-supervised learning (SSL) has been proven beneficial for mitigating the issue of limited labeled data, especially on volumetric medical image segmentation. Unlike previous SSL methods which focus on exploring highly confident pseudo-labels or developing consistency regularization schemes, our empirical findings suggest that differential decoder features emerge naturally when two decoders strive to generate consistent predictions. Based on the observation, we first analyze the treasure of discrepancy in learning towards consistency, under both pseudo-labeling and consistency regularization settings, and subsequently propose a novel SSL method called LeFeD, which learns the feature-level discrepancies obtained from two decoders, by feeding such information as feedback signals to the encoder. The core design of LeFeD is to enlarge the discrepancies by training differential decoders, and then learn from the differential features iteratively. We evaluate LeFeD against eight state-of-the-art (SOTA) methods on three public datasets. Experiments show LeFeD surpasses competitors without any bells and whistles, such as uncertainty estimation and strong constraints, as well as setting a new state of the art for semi-supervised medical image segmentation. Code has been released at https://github.com/maxwell0027/LeFeD.
- 1 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Medical Imaging
- Vol. 44 (1) , 44-56
Top Cited PapersConsistency-Guided Differential Decoding for Enhancing Semi-Supervised Medical Image Segmentation
Semi-supervised learning (SSL) has been proven beneficial for mitigating the issue of limited labeled data, especially on volumetric medical image segmentation. Unlike previous SSL methods which focus on exploring highly confident pseudo-labels or developing consistency regularization schemes, our empirical findings suggest that differential decoder features emerge naturally when two decoders strive to generate consistent predictions. Based on the observation, we first analyze the treasure of discrepancy in learning towards consistency, under both pseudo-labeling and consistency regularization settings, and subsequently propose a novel SSL method called LeFeD, which learns the feature-level discrepancies obtained from two decoders, by feeding such information as feedback signals to the encoder. The core design of LeFeD is to enlarge the discrepancies by training differential decoders, and then learn from the differential features iteratively. We evaluate LeFeD against eight state-of-the-art (SOTA) methods on three public datasets. Experiments show LeFeD surpasses competitors without any bells and whistles, such as uncertainty estimation and strong constraints, as well as setting a new state of the art for semi-supervised medical image segmentation. Code has been released at https://github.com/maxwell0027/LeFeD.
- Open Access
- 10 July 2024
- Published by Springer Nature in Visual Intelligence
- Vol. 2 (1) , 1-13
A human-in-the-loop method for pulmonary nodule detection in CT scans
Automated pulmonary nodule detection using computed tomography scans is vital in the early diagnosis of lung cancer. Although extensive well-performed methods have been proposed for this task, they suffer from the domain shift issue between training and test images. Unsupervised domain adaptation (UDA) methods provide a promising means to mitigate the domain variance; however, their performance is still limited since no target domain supervision is introduced. To make the pulmonary nodule detection algorithm more applicable in clinical practice and further boost the performance across domains, we propose a human-in-the-loop method in a semi-supervised fashion to enhance the model generalization ability when transferred from source domain to target domain. Specifically, we first train a detector model on source domain, and then the pre-trained detector is utilized with our proposed uncertainty-guided sample selection scheme (USSS) to find a few target domain samples worth annotating most and obtain their human annotations. Finally, the annotated and the rest unlabeled target domain samples are used together to refine the pre-trained model via our proposed zoom-in and zoom-out constraint (ZZC) strategy. We evaluate our method on the Nodule Analysis 2016 (LUNA16) and TianChi datasets. Experimental results show that our method surpasses recent competitive methods on source domain and also achieves surprising performance on target domain.
- 10 July 2024
- Published by Springer Nature in Visual Intelligence
- Vol. 2 (1) , 1-13
Open AccessA human-in-the-loop method for pulmonary nodule detection in CT scans
Automated pulmonary nodule detection using computed tomography scans is vital in the early diagnosis of lung cancer. Although extensive well-performed methods have been proposed for this task, they suffer from the domain shift issue between training and test images. Unsupervised domain adaptation (UDA) methods provide a promising means to mitigate the domain variance; however, their performance is still limited since no target domain supervision is introduced. To make the pulmonary nodule detection algorithm more applicable in clinical practice and further boost the performance across domains, we propose a human-in-the-loop method in a semi-supervised fashion to enhance the model generalization ability when transferred from source domain to target domain. Specifically, we first train a detector model on source domain, and then the pre-trained detector is utilized with our proposed uncertainty-guided sample selection scheme (USSS) to find a few target domain samples worth annotating most and obtain their human annotations. Finally, the annotated and the rest unlabeled target domain samples are used together to refine the pre-trained model via our proposed zoom-in and zoom-out constraint (ZZC) strategy. We evaluate our method on the Nodule Analysis 2016 (LUNA16) and TianChi datasets. Experimental results show that our method surpasses recent competitive methods on source domain and also achieves surprising performance on target domain.
- Open Access
Discrepancy Matters: Learning from Inconsistent Decoder Features for Consistent Semi-supervised Medical Image Segmentation
Semi-supervised learning (SSL) has been proven beneficial for mitigating the issue of limited labeled data especially on the task of volumetric medical image segmentation. Unlike previous SSL methods which focus on exploring highly confident pseudo-labels or developing consistency regularization schemes, our empirical findings suggest that inconsistent decoder features emerge naturally when two decoders strive to generate consistent predictions. Based on the observation, we first analyze the treasure of discrepancy in learning towards consistency, under both pseudo-labeling and consistency regularization settings, and subsequently propose a novel SSL method called LeFeD, which learns the feature-level discrepancy obtained from two decoders, by feeding the discrepancy as a feedback signal to the encoder. The core design of LeFeD is to enlarge the difference by training differentiated decoders, and then learn from the inconsistent information iteratively. We evaluate LeFeD against eight state-of-the-art (SOTA) methods on three public datasets. Experiments show LeFeD surpasses competitors without any bells and whistles such as uncertainty estimation and strong constraints, as well as setting a new state-of-the-art for semi-supervised medical image segmentation. Code is available at \textcolor{cyan}{https://github.com/maxwell0027/LeFeD}
Open AccessDiscrepancy Matters: Learning from Inconsistent Decoder Features for Consistent Semi-supervised Medical Image Segmentation
Semi-supervised learning (SSL) has been proven beneficial for mitigating the issue of limited labeled data especially on the task of volumetric medical image segmentation. Unlike previous SSL methods which focus on exploring highly confident pseudo-labels or developing consistency regularization schemes, our empirical findings suggest that inconsistent decoder features emerge naturally when two decoders strive to generate consistent predictions. Based on the observation, we first analyze the treasure of discrepancy in learning towards consistency, under both pseudo-labeling and consistency regularization settings, and subsequently propose a novel SSL method called LeFeD, which learns the feature-level discrepancy obtained from two decoders, by feeding the discrepancy as a feedback signal to the encoder. The core design of LeFeD is to enlarge the difference by training differentiated decoders, and then learn from the inconsistent information iteratively. We evaluate LeFeD against eight state-of-the-art (SOTA) methods on three public datasets. Experiments show LeFeD surpasses competitors without any bells and whistles such as uncertainty estimation and strong constraints, as well as setting a new state-of-the-art for semi-supervised medical image segmentation. Code is available at \textcolor{cyan}{https://github.com/maxwell0027/LeFeD}