We describe our methods for the HECKTOR 2025 challenge, which involved three tasks using PET/CT imaging and clinical data: segmentation of primary tumors and lymph nodes, recurrence-free survival prediction, and HPV-status classification. For tumor segmentation, we used a U-Net style SegResNet that achieved Dice scores of 0.52 for primary tumors and 0.38 for lymph nodes on the validation set, ranking in the Top 10. For survival prediction, we developed a multimodal model combining imaging features with clinical data, obtaining a C-index of 0.6482, also placing in the Top 10. The same framework applied to HPV-status classification yielded a balanced accuracy of 0.4655, earning 2nd place. Our results indicate that integrating features across data modalities improves performance, though class imbalance remains challenging. Future work could benefit from incorporating radiotherapy planning data and tumor volume measurements. Code is available at: https://github.com/BaixiangZ/hecktor2025
Accurate delineation of primary head and neck tumors and metastatic lymph nodes on PET/CT is critical for radiotherapy planning and prognostic assessment. Building on this clinical need, the HECKTOR 2025 challenge uses a large multi-centric dataset to provide a comprehensive benchmark for multimodal methods that integrate imaging and clinical information across three key tasks: segmentation of the primary tumor and involved lymph nodes, recurrence-free survival prediction, and HPV status classification. In this study, we (Team MEDAI) present our solutions for all three challenge tasks. For automated tumor segmentation, we employed an ensemble of ten lightweight STU-Net (small) models, achieving efficient and precise delineation of both primary tumors and metastatic lymph nodes. For recurrence-free survival prediction and HPV status classification, we developed a multimodal framework that integrates volumetric PET/CT imaging, lesion masks derived from the segmentation models, and structured clinical variables. Code is available at https://github.com/Liiiii2101/HECKTOR2025-MEDAI . Team: MEDAI.
Accurate segmentation of primary tumors (GTVp) and nodal lesions (GTVn) is essential for radiotherapy planning in head and neck cancer (HNC). The HECKTOR 2025 challenge provides a large-scale multi-center PET/CT dataset for benchmarking automated tumor segmentation. In this study, we developed a segmentation pipeline based on the nnU-Net v2 framework without architectural modifications. The model was trained using 5-fold cross-validation on the complete training cohort and deployed in a Docker container for submission. On the official validation leaderboard, our approach achieved a GTVp Dice of 0.7444, a GTVn Dice of 0.7956, and a GTVn F1-score of 0.5868, securing seventh place overall in Task 1. These results demonstrate that nnU-Net v2 remains a competitive baseline for multi-center PET/CT segmentation tasks, providing robust tumor delineation performance across heterogeneous datasets. We ranked third overall in the HECKTOR 2025 Challenge (Task 1).
Head and neck cancer poses major challenges for precision oncology, where accurate tumor segmentation and reliable survival prediction are essential yet remain difficult due to heterogeneous morphology and multi-center variability in PET/CT. We introduce HM-VNet, a unified multimodal framework designed to address these challenges through end-to-end learning. For segmentation, HM-VNet integrates hierarchical Transformer-based encoding with multimodal fusion to achieve robust delineation of both primary tumors and metastatic lymph nodes. For survival prediction, a deep cross-modal fusion network combines imaging, clinical, and radiomic features, further guided by anatomical priors derived from segmentation results. Comprehensive evaluation on the HECKTOR 2025 Challenge confirms that HM-VNet consistently outperforms state-of-the-art approaches, demonstrating strong effectiveness and promising clinical relevance in advancing automated multimodal intelligence for head and neck cancer management. Our source code is available at https://github.com/Wu-beining/HM-VNet . (Team: HDUMedAI)
This paper describes our submissions (team CDS) to the HECKTOR 2025 challenge, which addresses three tasks: (1) tumor and lymph node segmentation, (2) recurrence-free survival prediction, and (3) HPV status classification. For Task 1, we trained a baseline UNet and refined the final model using stochastic weight averaging and small lesion removal. For Task 2, we employed a lightweight 3D ResNet18 that combines PET, CT, segmentation masks, and clinical metadata, optimized with a Cox loss. For Task 3, we extended the segmentation model with a classification head and metadata integration. Cross-validation results were promising, performance on the preliminary validation set was however lower, underlining the challenges of generalization in multi-center cohorts. Code and trained models are available at github.com/JakobDexl/HECKTOR25 .
Head and Neck (H N) cancers represent one of the most prevalent malignancies worldwide, with accurate delineation of primary tumors and metastatic lymph nodes being essential for radiotherapy planning and treatment outcome prediction. Manual segmentation, though standard, is time-consuming and prone to inter-observer variability, motivating the development of automated deep learning–based methods. In this study, we propose HectoMixNet, a novel 3D encoder–decoder framework augmented with a bidirectional quasiseparable mixing module for robust segmentation of primary tumors and lymph nodes in multimodal PET/CT imaging. The proposed module efficiently integrates local anatomical detail with global volumetric context, addressing the challenges of heterogeneous and spatially complex lesions. We evaluate HectoMixNet on the HECKTOR 2025 Task 1 dataset, the largest multicentric benchmark for H N cancer analysis to date, comprising over 1,200 patients from 11 international centers. Our model achieved strong performance, with a GTVp Dice score of 0.8812, GTVn Dice of 0.8246, and GTVn F1 of 0.7273 on the validation and leaderboard sets. Compared to state-of-the-art baselines such as 3D Mamba and xLSTM, HectoMixNet demonstrated superior lymph node segmentation accuracy, highlighting the importance of bidirectional quasiseparable mixing for modeling long-range spatial dependencies. These results establish HectoMixNet as an effective and clinically relevant tool for automated H N cancer segmentation in large-scale, heterogeneous imaging cohorts.
We present a simple and effective pipeline for automatic detection and segmentation of primary tumors and lymph nodes in FDG-PET/CT for HECKTOR 2025 Task 1. The method starts with an anatomy-aware pre-crop of the head-and-neck region to suppress irrelevant context, followed by modality-specific intensity normalization with soft clamping. To mitigate cross-center domain shift, we apply single-subject, SSIM-guided spectrum swapping (SSIMH) on CT in the frequency domain without external references. For segmentation, we use a residual U-Net–style SegResNet with deep supervision and a combined Dice + Cross-Entropy loss. Training employs stratified five-fold cross-validation with foreground-centered sampling to emphasize small lesions. At inference, we use sliding-window tiling on the cropped volumes, lightweight post-processing to remove small isolated components, and a five-model ensemble by averaging per-voxel logits before softmax. On the official HECKTOR 2025 evaluation, our pipeline provides a concise and reproducible baseline for multi-center PET/CT tumor and lymph-node segmentation under domain shift. The overall recipe is concise and reproducible, providing a strong and transparent baseline for multi-center head-and-neck PET/CT segmentation under domain shift. (Team name: BIGS2)
Head and neck (H N) cancer segmentation from PET/CT is challenging due to heterogeneous imaging protocols across centers and the small proportion of tumor and lymph node volumes relative to the full field-of-view. We propose a two-stage coarse-to-fine framework for automatic segmentation of primary tumors (GTVp) and metastatic lymph nodes (GTVn) in the HECKTOR 2025 challenge. The framework first applies a head localization stage using an nnUNet to extract a coarse region-of-interest (ROI). In the fine segmentation stage, we integrate predictions from two complementary backbones: nnUNetResEncUNetLarge and MONAI-based SegResNet, both trained with five-fold cross-validation. To further enhance tumor delineation, especially on CT modality and across centers, we introduce multi-channel CT representations by concatenating raw CT, its squared intensity, cubic-root intensity, and PET as four input channels for nnUNetResEncUNetLarge. This design improves sensitivity to tumor intensity patterns and robustness against inter-center heterogeneity. The framework was evaluated quantitatively on the official test set of task 1 for the HECKTOR2025 challenge, achieving a GTVp DSC of 0.7341, a GTVn aggregated DSC of 0.7312, and a GTVn aggregated F1 score of 0.7260 as team SJTU_lab426.
Accurate prediction of human papillomavirus (HPV) status is essential for risk stratification and personalized treatment planning in head and neck cancer. In this work, we propose a multi-modal deep learning framework to classify HPV status using the HECKTOR25 Task 3 dataset, which provides 3D FDG-PET and CT scans with clinical data. Our approach leverages a 3D ResNet-18 architecture for imaging feature extraction, combined with a fully connected network to encode clinical variables, followed by multimodal fusion for final prediction. To address the significant class imbalance problem, we implemented a weighted cross-entropy loss. On internally held-out test splits, the model achieved a specificity of 0.9167 and a balanced accuracy of 0.9017, demonstrating robust intra-dataset performance. However, evaluation on the organizers external dataset—which contains cases from centers not included in the training data—yielded reduced performance (validation specificity 0.9048, balanced accuracy 0.6765), highlighting the challenges of cross-center generalization. These findings underscore the potential of multimodal deep learning for HPV status prediction and indicate that further strategies are required to enhance model robustness to inter-center variability.
The HECKTOR 2025 challenge provides a platform to benchmark automatic segmentation methods for Head and Neck (H N) primary tumors and lymph nodes in FDG-PET and CT scans (Task 1). This study presents a challenge submission based on a Residual Encoder U-Net within the nnU-Net framework, enhanced with modality-specific preprocessing and data augmentations. The proposed solution, submitted under the user name sebquet, achieved mean Dice scores of 71.81 https://github.com/sebquetin/Hecktor2025.git .
Purpose: Accurate segmentation and prognosis of head and neck cancer are crucial for effective treatment planning and personalized medicine. This study addresses two key challenges from the HECKTOR 2025 challenge: automated segmentation of primary gross tumor volume (GTVp) and prediction of Recurrence-Free Survival (RFS). Methods: For segmentation (Task 1), we employed the HecMamba architecture, leveraging its powerful HecMamab encoder to capture global context from PET/CT images. For prognosis (Task 2), we developed a multi-modal fusion model that combines a 3D ResNet for deep feature extraction from PET/CT images with a dedicated multi-layer perceptron (MLP) for processing clinical data. An ensemble of these models, trained using a 5-fold cross-validation strategy, was used to predict RFS. Results: Our segmentation model achieved a mean Dice Similarity Coefficient (DSC) of 0.785. The prognosis model achieved a high Concordance Index (C-index) of 0.902 on the test set, demonstrating strong predictive power by effectively integrating imaging and clinical features. Conclusion: This work presents a comprehensive deep learning framework that successfully addresses both segmentation and prognosis prediction for head and neck cancer. The HecMamba proves highly effective for segmentation, while our multi-modal fusion network demonstrates that integrating deep-learned imaging features with clinical data significantly enhances survival prediction accuracy.
Head and Neck Cancer (HNC) is a broad term for cancers that develop in the head and neck region. Accurate survival prediction is critical for guiding patient management and treatment planning. Traditional survival models, such as Kaplan–Meier curves and Cox Proportional Hazards models, are limited by their dependence on linearity and proportional hazards assumptions. Recently, deep learning–based survival models have demonstrated promising results for risk prediction. However, current approaches still struggle to fully integrate multimodal data and to capture region-specific features effectively. In this work, we employed a Multitask Learning framework that simultaneously performs Human Papillomavirus (HPV) classification, tumor segmentation, and survival prediction based on Positron Emission Tomography, Computed Tomography imaging, and clinical features. Integrating these three tasks into a unified model enables the use of shared feature representations. The segmentation module facilitates the extraction of tumor-specific features, while the classification branch incorporates HPV status prediction as a critical prognostic factor in HNC. This approach was developed and evaluated as part of our participation in the MICCAI 2025 HEad and neCK TumOR segmentation and outcome prediction challenge as the SIMS-LIFE team.
Head and Neck (H N) cancer is among the most common cancers worldwide, and its related clinical decision-making constitutes a systematic process that requires the integration of multimodal clinical data and the coordination of diverse tasks in the clinical workflow. However, how to effectively coordinate the interrelated clinical tasks to maximize their synergistic potential is still an open question. In this study, we propose a Multi-stage Multimodal Progressive Learning (named MMPL) framework for coordinated modeling of segmentation, diagnosis, and prognosis tasks, in the context of HECKTOR 2025 challenge at MICCAI 2025. Our MMPL progressively learns three clinical tasks that collectively facilitate personalized treatment planning: (i) tumor segmentation, (ii) HPV status classification, and (iii) survival prediction. Specifically, we establish a unified network backbone, consisting of a triple-stream encoder with adaptive PET/CT information fusion and an attention-gated decoder that can be applied to all three tasks. This backbone is successively trained for segmentation, classification, and survival prediction at three learning stages, where the knowledge is progressively learned with the guidance of prior knowledge accumulating from former stages. Further, the intermediate outputs (e.g., segmentation masks, HPV status) are leveraged as guidance on radiomics analysis or as supplementary indicators for the final prediction. In the final testing phase, our team InterStellar achieved top-tier performance across all three tasks, ranking first in HPV status classification and fourth in other two tasks.
Accurate delineation of gross tumor volumes (GTVs) in oropharyngeal carcinoma remains central to radiotherapy (RT) planning, and recent advances in automated segmentation are beginning to influence multi-ple aspects of clinical workflow. However, automated segmentation of head and neck primary tumors and nodal metastases in multi-modal FDG-PET/CT is challenging due to anatomical complexity and varying image resolution. We describe the DLaBella29 team’s approach for Task 1 of the MICCAI 2025 HECKTOR challenge, which involves a multi-phase deep learning pipeline for GTVp and GTVn segmentation. Our method leverages the MONAI Auto3DSeg framework with a 3D SegResNet backbone, trained on co-registered PET/CT scans using a cross-validation strategy (3/7/15 folds across Phases). Key innova-tions include a mandible-anchored region-of-interest cropping strategy derived from automated mandible segmentation to focus the model on the oropharyngeal region and improve efficiency. A multi-phase seg-mentation pipeline was employed: an initial Phase infers GTVp and GTVn in the focused anatomical region-of-interest, and a second Phase independently refines these predictions. Post-processing ensues to merge GTVp and GTVn outputs to the original image coordinates. On the HECKTOR 2025 test set, our more limited and resource-constrained algorithm achieved a primary tumor Dice Similarity Coefficient (DSC) of 0.5289, an aggregated nodal DSC of 0.6156, and a GTVn detection F1-score of 0.5561, indicating fair performance in the challenge. This paper details the clinical context, methodology, cross-validation and testing results, and the implications of PET/CT-guided automated segmentation for oropharyngeal carcinoma.
Accurate segmentation of Head and Neck (H N) tumor and lymph node in PET/CT images is essential for diagnosis and treatment planning. However, it remains challenging due to heterogeneous lesion morphology and variable physiological uptake. In this paper, we propose a 3D U-Net based architecture with a PET-guided Spatial Attention Module (PSAM). PSAM consumes the PET image to generate a spatial attention map that gates encoder skip features. Furthermore, we use Squeeze-and-Excitation (SE) Normalization, which dynamically recalibrates channel responses and improves multimodal fusion. Notably, we introduce a novel molecular-information-guided preprocessing pipeline, which reduces the input volume size. This design enables modality-aware modeling and aims for robust generalizable segmentation of both primary Gross Tumor Volume (GTVp) and nodal Gross Tumor Volume (GTVn) across multi-center PET/CT data. We obtain a mean Dice score of 0.6073 with class-wise Dice scores 0.7133 (GTVp) and 0.5013 (GTVn) on the validation set. On the official test set, our model achieved a Dice score of 0.3803 (GTVp) and 0.3733 (GTVn). Further studies are required to improve generalization performance across multi-center datasets.
Atypical mitosis detection remains a critical task in computational pathology, offering insights into tumor aggressiveness and diagnostic reliability. In the MIDOG 2025 Track 2 challenge [1], we propose the Dual-Fusion Double-Ensemble (DFDE) framework, which integrates multi-model feature fusion and ensemble learning. Precomputed features are obtained by concatenating HIBOU-L [2] and Barlow Twins [3] embeddings. Classification is performed using Linear Discriminant Analysis (LDA) followed by CatBoost with focal loss [4] and XGBoost [5]. Final predictions are derived through a weighted soft-voting ensemble. On the official preliminary test set, DFDE achieved a Balanced Accuracy of 0.7012, with sensitivity of 0.6620, specificity of 0.7405, and ROC AUC of 0.7475. These results highlight the importance of feature fusion and ensemble methods for domain-generalized mitosis classification. The full implementation of DFDE is publicly available at https://github.com/DigitalPathology/DFDE-MIDOG2025 .
Accurate detection and classification of mitotic figures in histopathology remains difficult due to their infrequent occurrence, morphological variability, and strong domain shifts across scanners, species, and tumor types. The Mitosis Domain Generalization (MIDOG) Challenge 2025 addresses these limitations through two complementary tasks: robust mitotic figure detection across different tumor regions (Track 1) and atypical vs. normal mitotic figure classification (Track 2). In this paper, we provide the official baseline methods for both tracks: a Fully Convolutional One-Stage (FCOS)-based detector for Track 1 and an EfficientNet-V2 classifier for Track 2. Both tasks use the multi-center, multi-species MIDOG++ training dataset and are evaluated on a diverse test set containing 12 tumor types, including domains absent from training, and three region-of-interest categories (hotspot, random, potentially challenging). The detection baseline achieved an overall F1 score of 0.6883 on the final test set, with strong performance in hotspot regions (F1 = 0.7472) but reduced robustness in random (F1 = 0.6500) and challenging (F1 = 0.4161) region of interests (RoIs). The classification baseline reached a balanced accuracy of 0.8274 and ROC AUC of 0.9067. These baselines offer a reproducible reference for the participants in the MIDOG 2025 challenge.
The MIDOG 2025 challenge extends prior work on mitotic figure detection by introducing a new Track 2 on atypical mitosis classification. This task aims to distinguish normal from atypical mitotic figures in histopathology images, a clinically relevant but highly imbalanced and cross-domain problem. We investigated two complementary backbones: (i) ConvNeXt-Small, pretrained on ImageNet, and (ii) a histopathology-specific ViT from Lunit trained via self-supervision. To address the strong prevalence imbalance (9,408 normal vs. 1,741 atypical), we synthesized additional atypical examples to approximate class balance and compared models trained with real-only vs. real+synthetic data. Using five-fold cross-validation, both backbones reached strong performance (mean AUROC ≈ 95
The reliable differentiation of normal mitotic figures (MFs) and atypical MFs is a challenging yet clinically relevant task in cancer histopathology. While recent deep learning (DL) approaches have substantially improved automated MF detection and classification, they still struggle with subtle morphological variations and inter-class similarities. In this work, a hybrid classification framework is introduced that integrates a deep ensemble of three ConvNeXt-Base models with a complementary Rule-Based Refinement (RBR) module. All models were trained on a harmonized dataset comprising AMi-Br, MIDOG25, and OMG-Octo Atypical, totaling 15,689 mitotic figures. The DL classifier (DLC) incorporates stain normalization, extensive augmentation, and transfer learning to achieve robust feature extraction, reaching a balanced accuracy of 89.18
Accurate identification of atypical (AMF) versus normal mitotic figures (NMF) is crucial, as cellular composition and mitotic activity are key determinants for cancer grading and prognosis. This paper presents our contribution to the MIDOG 2025 Challenge, introducing an efficient approach to classify AMFs from NMFs. Classification of atypical versus normal mitotic figures is challenging due to subtle morphological differences, high intra-class variability, and domain shifts across organs, tissues, and scanners. In this work, we fine-tuned a foundation model Virchow2, pretrained on histopathology images, using Low-Rank Adaptation (LoRA), training only approximately ≈ 5.27 million parameters. This was combined with extensive augmentation, weighted sampling, and Focal Loss to address class imbalance and domain heterogeneity. Test-time augmentation (TTA) is applied to further enhance model robustness. Our best-performing model achieved a balanced accuracy of 90.07
