From Infusion to Assimilation Distillation for Medical Image Segmentation

Junsong Yuan, Jiankang Hong, Ye Luo, Yinan Liu

cvpr2026-06-01

Analysis: not analyzed

Local PDF: stored

Source PDF: linked

Source PDF

Research metadataShow details

Affiliations: Not available
Published: 2026-06-01
Processed: 6/20/2026, 6:33:26 PM
Analysis model: Not available
Analysis status: not_analyzed
Local PDF artifact: papers/pdf/2026/from-infusion-to-assimilation-distillation-for-medical-image.pdf

Abstract

Although foundation models (e.g. SAM) perform remarkably in medical image segmentation, its high computational complexity limits deployment. Knowledge distillation (KD) allows lightweight models to inherit the representational capabilities of large models, thereby mitigating this issue. Existing KD methods enhance student performance, but due to teacher-student different feature advantages, they neglect to internalize and integrate student's semantic information adaptively after knowledge transfer, causing poor knowledge assimilation and limiting gains and generalization. To address this limitation, we propose a novel medical image segmentation framework, which is infusion to assimilation distillation (IAD). In Knowledge Infusion Stage (KIS), to semantically align teacher-student prediction distributions, soft-label distillation is combined with class-weighted prototype alignment strategy. In Knowledge Assimilation Stage (KAS), to promote adaptively semantic assimilation, a contrastive semantic self-optimization strategy refines student predictions through positive and negative sample pairs and imposes reverse constraints on encoder features to enhance semantic consistency. IAD achieves DICE gains of 4.32% on Synapse, 1.85% on ACDC, and 2.42% on Polyp datasets, and delivers an average 4.16% generalization gain on ISIC2018, PH2, BUSI, and STU datasets, outperforming mainstream KD methods. Code is available at https://github.com/hjklearn/IAD.

Stored markdown is an ingestion scaffold, so it remains excluded until substantive methodology, result, and limitation sections are available.

Analysis unavailable

Only an ingestion scaffold is stored; substantive analysis is not available yet.

Additional figures

Figure 1. A performance drop (i.e. DICE and mIoU) was ob- served in 58% of the student models on transfer datasets across 12 mainstream Knowledge Distillation methods, while the rest ones achieve relatively limited improvement compared to our IAD model. The y-axis denotes typical KD methods investigated.

Figure 2. Visualization of the unexpected student model’s performance degradation in typical KD methods on Synapse dataset. (a) ground truth; (b) and (c) are the predictions of the baseline teacher and student models, respectively; Predictions of (d) CILD [52] and (e) LSKD [42]; (f) our IAD , selectively transferring teacher knowledge and preserving student’s valuable capabilities, achieves best result. that most methods’ student models exhibit limited perfor- mance gains (e.g. VL2Lite [24], DIST+ [22], LSKD [42],

Figure 3. The architecture of proposed IAD. In stage1, we first inject knowledge via soft-label distillation with class-weighted proto- type alignment to transfer class-level semantics and alleviate class imbalance. In stage2, we then perform knowledge assimilation using a contrastive semantic self-optimization strategy, which refines prediction semantics through positive-negative relational modeling and reinforces feature consistency via reverse constraints, enabling adaptive semantic assimilation. tion to address logit magnitude dependency. CrossKD [48]

Figure 4. The t-SNE visualization of student per-class proto- types on Synapse dataset after adding KIS and KAS sequentially.

Figure 5. The t-SNE of KAS-induced class position reordering.

Figure 6. Visualization comparison of our method and repre- sentative KD baselines on Synapse [28] dataset.

Figure 7. Training curves of DICE over epochs. Black curve represents our method. Black dot indicates peak DICE value.

Figure 8. Ablation visuals of IAD on Synapse [28] dataset. ISIC2018 and BUSI datasets. As shown in Fig. 7, through- out training, our method consistently achieves higher DICE scores than most KD methods, reaching the peak value, demonstrating superior stability and efficiency. To further

Figure 9. Sensitive analysis of class weight ratios on Synapse. Table 8. Ablation study of KD losses on Synapse [28] dataset. Number 1

Figure 10. Visual results of KD methods before and after KAS. Table 10. Effect of our IAD on diverse student architectures.