Early prediction of post-surgical recurrence in Non-Small Cell Lung Cancer (NSCLC) is crucial for improving patient outcomes. Radiomics, particularly multimodal approaches, shows promise in enhancing predictive accuracy. However, studies using PET, CT, and clinicopathological (CP) data to improve prediction remain limited. Notably, the integration of these modalities through ensemble stacking for decision-level fusion has yet to be fully explored.This study evaluates radiomics extracted from positron emission tomography (PET) and computed tomography (CT) scans-individually, combined, and integrated with CP data- for NSCLC recurrence classification. A cohort of 131 patients underwent PET and CT imaging, with CP variables collected from The Cancer Imaging Archive (TCIA). Radiomics features were extracted by pyradiomics library. Models were developed through feature concatenation followed by decision fusion using ensemble stacking and were assessed using precision, recall, F1 score, accuracy, and AUC.Results show PET+CT fusion achieved the highest performance (AUC = 0.80), while CP integration did not improve the performance and, in some cases, negatively affected the performance results. These findings suggest that optimized radiomics models alone may suffice for predictive modeling in NSCLC.Clinical Relevance-The robust performance of PET+CT models highlights the potential of non-invasive radiomics for personalized recurrence assessment in NSCLC. By integrating these models into clinical workflows, clinicians can tailor followup strategies and treatment plans, ensuring patients receive neither excessive nor insufficient therapy after surgery. This approach optimizes care, enhances surveillance, and improves patient outcomes.

Digital biomarkers offer novel approaches to non-invasive health monitoring, particularly in Palliative Care, for example in cancer patients, where the identification and relief of symptom burden and distress are the leading goals of care. This study investigates the correlation between acoustic speech features and distress severity in female cancer patients, using the Edmonton Symptom Assessment System (ESAS) as a reference. Speech recordings were collected from 28 cancer patients at up to four different time points, with acoustic features extracted using the openSMILE toolkit (ComParE 2016 feature set). The analysis focused on the 23rd and 24th Mel filter-bank bands-MFB 23 and MFB 24-which are two individual channels of the 26-channel Mel filter-bank computed by openSMILE. Pearson correlation analysis identified 13 spectral features significantly associated with the ESAS score of the female cohort. A multivariate Ordinary Least Squares (OLS) regression model demonstrated that selected acoustic parameters explained 29% of the variance in distress levels, with spectral flatness and mean energy in MFB 23 emerging as key predictors. These findings suggest that speech-based biomarkers may facilitate automated, objective distress screening in oncology patients. By integrating acoustic analysis into clinical workflows, this study highlights the potential of digital voice biomarkers for continuous symptom monitoring. Future research should refine predictive models and expand patient cohorts to enhance clinical applicability.Clinical relevance-This study highlights the potential of speech-derived digital biomarkers for distress screening in female palliative oncology patients. By correlating acoustic speech features with ESAS scores, it demonstrates a noninvasive, objective method for symptom monitoring. Integrating voice analysis into clinical workflows could enhance early intervention, reduce patient burden, and improve precision in symptom management, for example by integration into telemedicine interventions and follow-ups.

The spread of early-stage adenocarcinomas to locoregional lymph nodes is a critical event in disease progression of gastric cancer. Multiple instance learning (MIL) is widely employed in computational pathology to solve the absence of pixel-wise or patch-wise annotations in Whole Slide Image (WSI) datasets. MIL algorithms are typically applied a single-scale of WSIs, whereas pathologists usually aggregate diagnostic information across multiple scales. To this end, we propose a novel cross-scale MIL framework with multi-scale interactions to predict lymph node metastasis (LNM) from early-stage gastric cancer (EGC) WSIs. A novel cross-scale attention module is proposed to obtain cross-scale features from different resolutions with multi-scale interaction. Cross-Scale features, along with resolution-specific features, are then aggregated for the final slide-level prediction. Our experiments are conducted on a clinical cohort of WSIs from 740 patients with T1-stage gastric cancer. Our approach achieves a superior Area under the Curve (AUC) of 0.712, outperforming baseline MIL models. Additionally, multi-scale attention visualizations are generated to enhance the interpretability of automatic LNM diagnosis.Clinical Relevance- This study develops a deep learning model for LNM prediction from EGC WSIs, as opposed to lymph node specimens. Our research would be helpful to make treatment decisions for EGC patients, for example, avoid unnecessary lymph node resection.

Synthetic lethality (SL) holds great promise as an emerging strategy of cancer therapy by selectively eliminating cancer cells. Despite wide adoption of high-throughput technologies for SL screening, the limited understanding of SL mechanisms poses significant challenges to its clinical application. Consequently, computational methods for uncovering SL mechanisms are of considerable value. In recent years, the widespread adoption and rapid advancement of large language models (LLMs) have made artificial intelligence-driven explanations of SL mechanisms increasingly feasible and reliable. In this paper, we propose a novel SL mechanism explanation framework based on LLMs that integrates and complements GraphRAG (Graph Retrieval-Augmented Generation) with knowledge graphs. Our approach named SL-MERK combines the SL interaction patterns extracted by GraphRAG from the biomedical literature with the rich mechanistic knowledge encoded in knowledge graphs. Leveraging the capabilities of LLMs for generalization and natural language generation, this framework generates comprehensive and interpretable natural language explanations of SL mechanisms. Furthermore, experimental evaluations demonstrate that our framework significantly outperforms GPT-4 and several other LLMs in explanatory performance.

Head and neck radiotherapy often combines a patient's MRI, showing soft tissue contrast, with a pre-treatment CT allowing for dosimetry planning. Synthesizing missing CT data from available MR images minimizes radiation exposure, and facilitates adaptive re-planning. We propose a generative diffusion model that synthesizes CT images of tumors from available MR modalities, incorporating structural guidance within a stacked diffusion framework. The model utilizes two stacked denoising diffusion probabilistic models (DDPMs). The first is a structure image generator, producing structural representations of CT images from the corresponding MRI inputs. These representations are then utilized by a second contextual image DDPM, which leverages both the original MRI and the generated structural representations as an augmented multi-channel input to improve the synthesis of the CT images. Our training employs a variational inference approach that combines a lower variational bound loss with a mean absolute error loss, leveraging both structural and contextual features. Evaluated on the Head and Neck Organ-at-Risk Multi-Modal dataset (HaN-Seg), our model outperforms recent MR-to-CT generative diffusion models, achieving a multiscale structure similarity index (multiscale-SSIM) of 0.85 ± 0.08, a mean absolute error (MAE) of 0.09 ± 0.06, and a peak signal-to-noise ratio (PSNR) of 22.05 ± 1.83. Additionally, the model achieved the highest probability rand index (PRI) score of 0.83 ± 0.04 with a Dice score of 0.75 ± 0.07, and a global consistency error (GCE) of 0.16 ± 0.05 on segmented tumor area of synthetic sCT images.

Lung cancer is one of the leading causes of cancer-related deaths worldwide, mainly due to late diagnosis. Screening programs can benefit from Computer-Aided Diagnosis (CAD) systems that detect and classify lung nodules using Computed Tomography (CT) scans. A great proportion of the literature proposes deep learning models based on single and private datasets with no evaluation of their generalisation capability. The main goal of this work is to study and address the lack of generalisation to out-of-domain data (source domain different from the target domain). In this work, we propose using a ResNet architecture with 2.5D inputs capable of maintaining the spatial information of the nodules (3 input channels based on the anatomical planes). Secondly, we apply domain-specific data augmentation tailored for CT scans. Combined with data augmentation, using 2.5D inputs achieves the best results, both in in-domain data (LIDC-IDRI: N=1377 nodules; and LNDb: N=183 nodules) and in out-of-domain data (LUNGx: N=73 nodules). In in-domain data, an Area Under the Curve (AUC) of 0.914 was achieved in the internal test set and 0.746 in one of the external test sets. Notably, in out-of-domain data, where the ground-truth labels have been confirmed by biopsy, whereas the training data only involved radiologist annotation regarding the "likelihood of malignancy", AUC improves from 0.576 to 0.695, reaching a performance close to that of radiology experts. In the future, strategies should be applied to deal with the level of uncertainty of lung nodule annotations based solely on the observation of the CT scans.Clinical relevance- This work provides an automatic method for lung nodule malignancy classification based on CT scans, combined with generalisation methods that allow a good performance across different cohort populations and hospitals.

Accurate colorectal polyp segmentation is crucial for the early detection and prevention of colorectal cancer, one of the leading causes of cancer-related deaths worldwide. While colonoscopy remains the most reliable screening method, it is time-consuming, resource-intensive, and highly dependent on the operator, which can lead to variability in diagnosis and potential delays. Deep learning models have shown great potential in automating polyp detection, but their large size and high computational demands make them impractical for real-time clinical use. To overcome these challenges, we introduce LPD-Net, a lightweight and efficient alternative to DUCK-Net that reduces computational complexity while maintaining high segmentation accuracy. This is achieved by optimizing the network architecture, reducing the number of residual blocks, and leveraging depthwise and pointwise convolutions. Our model strikes a balance between performance and computational efficiency. With robust preprocessing and test-time augmentation, LPD-Net achieves state-of-the-art segmentation on CVC-ClinicDB and Kvasir-SEG while remaining lightweight.Clinical RelevanceEarly and precise polyp segmentation is essential for effective colorectal cancer screening and treatment. LPD-Net ensures high segmentation accuracy while significantly reducing parameters, enabling real-time analysis of colonoscopy images. Its lightweight design lowers computational costs, making it suitable for resource-limited settings. By enhancing segmentation efficiency and robustness, LPD-Net supports faster and more reliable polyp assessment, aiding timely medical intervention and improved patient outcomes.

Papanicolaou (Pap) stain is used to stain cells to visually assess lesions or determine if they are benign or malignant in cytological examinations. In digital cytology imaging, Pap stain unmixing has been proposed to enable objective interpretation of stain abundance for the diagnosis based on quantitative measurement. Using sparsity-based regularization, our previously proposed method allowed stain unmixing for RGB images with fewer channels than Pap staining dyes. Its effectiveness was demonstrated with simulated RGB images converted from multispectral data. In this study, we validate the robustness of stain quantification against color variations induced by different imaging systems, proving this method works in actual RGB systems. The robustness is achieved by extracting the stain matrix from single stain images and normalizing with the robust maximum values of stain abundance. Additionally, we apply this technique to classify cytoplasmic mucin in lobular endocervical glandular hyperplasia and normal endocervical cells. The results yield high accuracy while allowing the colors to be expressed quantitatively through the amounts of dyes, which can serve as the criterion for judgment. The cell classification is achieved using RGB images obtained from a practical whole slide image scanner, in which the classifier is trained with RGB images generated from multispectral data.Clinical Relevance-This establishes the efficacy of Papanicolaou stain unmixing for the robust detection of lobular endocervical glandular hyperplasia cells in RGB images obtained from various imaging systems.

We introduce a novel approach for distinguishing between benign and malignant tumors in breast ultrasound images using a set of features derived from third-order statistics. Unlike second-order statistics, which measure relationships between pairs of pixels, third-order statistics capture correlations among three pixels (triple correlation), providing a complete characterization of the information within an image. The third-order features were computed from the triple correlation distribution and include the mean, standard deviation, kurtosis, skewness, and entropy. Our findings reveal that third-order features significantly outperform commonly used second-order ones in key classification metrics, including accuracy, F1 score, specificity, and precision in random forest classification. These results suggest that metrics based on third-order statistics can uncover previously unrecognized patterns, offering valuable insights for improving tumor classification in ultrasound imaging.Clinical Relevance- This study demonstrates the enhanced classification capability of third-order features compared to second-order features for breast ultrasound tumors. We propose that the novel integration of these metrics can improve classification performance and can serve as valuable additions to the toolkit for breast ultrasound analysis, particularly for borderline or questionable tumor pathology. Furthermore, the versatility of this approach extends to other modalities including mammography and MRI.

Thyroid cancer, a prevalent malignancy with rising incidence-especially among females-requires early detection to reduce overdiagnosis risks and enable less invasive treatments. Ultrasound (US) imaging is pivotal but limited by observer variability. We propose a 3D imaging system combining a US machine, robotic arm, depth camera, and customized software to address this. The system integrates automated robotic assistance for precision and repeatability, alongside 3D reconstruction of ultrasound elastography (USE) and brightness mode (USB) for flexible multi-view observation and quantitative stiffness analysis. Our system incorporates SegUnXt+, a deep learning model demonstrating competitive thyroid gland segmentation (IoU 82.8%, DC 90.6%) and thyroid nodule segmentation (IoU 71.9%, DC 83.3%), outperforming other models. The system enhances diagnostic accuracy by minimizing observer dependency, enabling precise 3D thyroid visualization, and supporting early detection through automated, quantitative elastographic and morphologic analysis.Clinical Relevance- Mass screening and support diagnosis of thyroid cancer to improve screening accuracy and accessibility.