Jingping Liu, Xingchen Peng, Yan Zhou, Ziyan Liu, Jie Zhai, Ronghao Chen, Huacan Wang, Xiaofeng Jia. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Survival prediction is crucial for cancer patients as it provides essential early prognostic information for treatment planning and decision making. Despite impressive performance , current multi-modal survival prediction methods that integrate pathology and genomic data face two main challenges: (1) Whole-slide images (WSIs) generally exhibit hierarchical structures, but the interactions of phenotypes at different resolutions remain unexplored. More importantly, the potential semantic discrepancy arising from diverse resolutions is often ignored. (2) The absence of effective interactions between the inherent hierarchical structures of WSIs and genomic data. To address these challenges, in this paper, we propose Cross-modal Dual-interaction Transformer (CD-Former), a robust hierarchical framework for multi-modal survival prediction. Our CD-Former involves two key components: (1) an Multimodal Cross-Scale Calibration (MCSC) module for effectively capturing correlations across multiple resolutions and calibrating fine-grained features, thereby bridging the semantic discrepancy caused by different WSI resolutions; and (2) a hierarchical interaction module termed Multi-modal Dual-interaction (M2Di) for fully exploring multi-resolution cross-modal correlations and interactions, which comprises a Patch-level Cross-Attention Block (PCAB) and a Region-level Cross-Attention Block (RCAB) to investigate cross-modal associations between patch- or region-level features of WSI and genomic data. Additionally, we employ a scale-oriented WSI enhancer to capture the interactions among various components of WSIs. The experimental results demonstrate the effectiveness of our proposed framework, which achieves state-of-the-art performance compared to previous studies.

Accurate dose distribution prediction is crucial in the radiotherapy planning. Although previous methods based on convolutional neural network have shown promising performance, they have the problem of over-smoothing, leading to prediction without important high-frequency details. Recently, diffusion model has achieved great success in computer vision, which excels in generating images with more high-frequency details, yet suffers from time-consuming and extensive computational resource consumption. To alleviate these problems, we propose Frequency-Decomposed Diffusion Model (FDDM) that refines the high-frequency subbands of the dose map. To be specific, we design a Coarse Dose Prediction Module (CDPM) to first predict a coarse dose map and then utilize 2D discrete wavelet transform to decompose the coarse dose map into a low-frequency subband and three high-frequency subbands. There is a notable difference between the coarse predicted results and ground truth in high-frequency subbands. Therefore, we design a diffusion-based module called High-Frequency Refinement Module (HFRM) that performs diffusion operation in the high-frequency components of the dose map instead of the original dose map. Extensive experiments on two in-house datasets verify the effectiveness of our approach.

Radiotherapy is one of the leading treatments for cancer. To accelerate the implementation of radiotherapy in clinic, various deep learning-based methods have been developed for automatic dose prediction. However, the effectiveness of these methods heavily relies on the availability of a substantial amount of data with labels, i.e. the dose distribution maps, which cost dosimetrists considerable time and effort to acquire. For cancers of low-incidence, such as cervical cancer, it is often a luxury to collect an adequate amount of labeled data to train a well-performing deep learning (DL) model. To mitigate this problem, in this paper, we resort to the unsupervised domain adaptation (UDA) strategy to achieve accurate dose prediction for cervical cancer (target domain) by leveraging the well-labeled high-incidence rectal cancer (source domain). Specifically, we introduce the cross-attention mechanism to learn the domain-invariant features and develop a cross-attention transformer-based encoder to align the two different cancer domains. Meanwhile, to preserve the target-specific knowledge, we employ multiple domain classifiers to enforce the network to extract more discriminative target features. In addition, we employ two independent convolutional neural network (CNN) decoders to compensate for the lack of spatial inductive bias in the pure transformer and generate accurate dose maps for both domains. Furthermore, to enhance the performance, two additional losses, i.e. a knowledge distillation loss (KDL) and a domain classification loss (DCL), are incorporated to transfer the domain-invariant features while preserving domain-specific information. Experimental results on a rectal cancer dataset and a cervical cancer dataset have demonstrated that our method achieves the best quantitative results with [Formula: see text], [Formula: see text], and HI of 1.446, 1.231, and 0.082, respectively, and outperforms other methods in terms of qualitative assessment.

Radiation therapy is a fundamental cancer treatment in the clinic. However, to satisfy the clinical requirements, radiologists have to iteratively adjust the radiotherapy plan based on experience, causing it extremely subjective and time-consuming to obtain a clinically acceptable plan. To this end, we introduce a transformer-embedded multi-task dose prediction (TransMTDP) network to automatically predict the dose distribution in radiotherapy. Specifically, to achieve more stable and accurate dose predictions, three highly correlated tasks are included in our TransMTDP network, i.e. a main dose prediction task to provide each pixel with a fine-grained dose value, an auxiliary isodose lines prediction task to produce coarse-grained dose ranges, and an auxiliary gradient prediction task to learn subtle gradient information such as radiation patterns and edges in the dose maps. The three correlated tasks are integrated through a shared encoder, following the multi-task learning strategy. To strengthen the connection of the output layers for different tasks, we further use two additional constraints, i.e. isodose consistency loss and gradient consistency loss, to reinforce the match between the dose distribution features generated by the auxiliary tasks and the main task. Additionally, considering many organs in the human body are symmetrical and the dose maps present abundant global features, we embed the transformer into our framework to capture the long-range dependencies of the dose maps. Evaluated on an in-house rectum cancer dataset and a public head and neck cancer dataset, our method gains superior performance compared with the state-of-the-art ones. Code is available at https://github.com/luuuwen/TransMTDP.