Computational Mathematics Modeling in Cancer Analysis

Multimodal learning, integrating histology images and genomics, promises to enhance precision oncology with comprehensive views at microscopic and molecular levels. However, existing methods may not sufficiently model the shared or complementary information for more effective integration. In this study, we introduce a Unified Modeling Enhanced Multimodal Learning (UMEML) framework that employs a hierarchical attention structure to effectively leverage shared and complementary features of both modalities of histology and genomics. Specifically, to mitigate unimodal bias from modality imbalance, we utilize a query-based cross-attention mechanism for prototype clustering in the pathology encoder. Our prototype assignment and modularity strategy are designed to align shared features and minimizes modality gaps. An additional registration mechanism with learnable tokens is introduced to enhance cross-modal feature integration and robustness in multimodal unified modeling. Our experiments demonstrate that our method surpasses previous state-of-the-art approaches in glioma diagnosis and prognosis tasks, underscoring its superiority in precision neuro-Oncology.

Laparoscopic exploration of the abdominal cavity is routinely performed for the diagnosis, assessment, and staging of peritoneal metastasis (PM). Accurately measuring tumor size during this procedure is crucial for prognosis and treatment planning. As conventional approaches for tumor size measurement rely on subjective manual assessments during or after surgery, they stand to benefit from computer assistance. This study proposes a new method for measuring tumor size in laparoscopic monocular videos. Specifically, we introduce a novel mathematical equation that connects the intrinsic parameters of a monocular camera, the surface area of target and reference objects, and their distances to the camera. Furthermore, we combine this equation with an object segmentation model (Mask2Former) and a depth estimation model (MiDaS), creating an end-to-end framework that automates tumor size measurement in monocular laparoscopic videos. We evaluate the proposed method using a laparoscopy dataset comprising 18 videos depicting 76 tumor biopsies, with tumor size measured by surgeons who are experts in laparoscopic surgery. When estimating the size of the various tumors in this dataset, we obtain a Mean Absolute Error (MAE) of 2.44 mm ± 0.23 mm, demonstrating that the newly proposed method accurately predicts intraoperative tumor size. Our code and the evaluation dataset are publicly available on https://​github.​com/​amiiiirrrr/​TSEMLV.

Follicular lymphoma (FL) is a non-Hodgkin lymphoma and an indolent B-cell lymphoproliferative disorder of transformed follicular center B cells. In the diagnosis, FL should be graded by counting the number of centroblasts in the pathological image, which is time-consuming. In this study, we try to propose a FL grading method based on the PET and CT images. We propose a 3D-DDcGAN to fuse the simultaneously collected PET and CT images. Then, the BayesianResNet18 (ResNet18 is improved by introducing Bayes’ theorem) is adopted for the FL grading. Our method is trained and tested on mixed data consisting of FL grades I-III and DLBCL. Finally, the evaluation metrics for our method are accuracy 0.814, precision 0.782, recall 0.699, macro-averaged F1-score 0.731, and micro-averaged F1-score 0.817. The method based on deep learning and medical imaging will help assist in disease grading and developing personalized treatment plans.

The accurate identification of circulating abnormal cells (CACs) in four-color fluorescence images is highly dependent on the fluorescence expression under each channel. Previous studies have utilized instance segmentation and target detection algorithms to identify cells and signal points in four-color fluorescence in situ hybridization (FISH) microscopy images. However, these algorithms require high accuracy in cell edge segmentation and signal point detection, which hinders the success of CAC detection. In this study, we propose a novel method for discriminating CAC cells using four-color fluorescence channels and the fusion of information from different models based on previous techniques. In particular, we utilize the information distribution of the four-color fluorescence channels to train the MMFM-DL network and MMFM-ML model, respectively. Thereafter, the model fusion strategy is employed to enhance the performance of the deep learning and shallow machine learning methods, thereby achieving a more comprehensive and accurate identification of CACs. This method requires the interpretation of only 0.92% of the cells in the microscopy images of four-color fluorescence in situ hybridization, thereby ensuring that the CAC recall rate is guaranteed to be over 98%. This is a significant improvement over the previous method.

Single domain generalization aims to address the challenge of out-of-distribution generalization problem with only one source domain available. Feature disentanglement is a classic solution to this purpose, where the extracted task-related feature is presumed to be resilient to domain shift. However, the absence of references from other domains in a single-domain scenario poses significant uncertainty in feature disentanglement (ill-posedness). In this paper, we propose a new framework, named Domain Game, to perform feature disentangling for medical image segmentation, based on the observation that anatomical features are more sensitive to geometric transformations, whilst domain-specific features probably will remain invariant to such operations. Results from cross-site test domain evaluation showcase approximately an \(\sim \)11.8% performance boost in prostate segmentation and around \(\sim \)10.5% in brain tumor segmentation. The codes will be available at https://​github.​com/​chqwer2/​Domain-Game.

The multi-omics integration gives a whole new perspective into pathway analysis to reveal the complicated nature of cellular systems. While the understanding of interactions among different omics data remains unknown, current methods do not consider the unique and similar properties. In this paper, we propose Attention-fusion Model for Multi-Omics (AMMO), a robust method that addresses this challenge through domain separation. Our proposed attention-based approach inherently captures the similarities and differences across various omics modalities, enhancing the interpretability of the integrated data. Our proposed method can achieve a state-of-the-art C-index of 0.8017 in overall survival prediction in TCGA-LUAD data with the diverse types of omics data: DNA Methylation, exon expression RNA Seq (HiC), and protein expression (RPPA). We also demonstrated the performance increase by adding more modalities with the ablation test, the results confirmed our assumption of improving model performance by including more modalities to our method.

Immune checkpoint inhibitors (ICIs) have revolutionized the treatment of non-small cell lung cancer (NSCLC), benefiting 20–30% of patients. The current clinical standard for initiating ICI therapy is the assessment of Programmed Death-Ligand 1 (PD-L1) status via immunohistochemistry (IHC) on biopsy specimens. However, this invasive procedure presents risks and limitations, highlighting the need for a non-invasive alternative. This study retrospectively analyzed a cohort of 746 patients with stage IV metastatic NSCLC undergoing immunotherapy, divided into training (n = 298), internal validation (n = 75), and testing (n = 360) groups. Thirteen cases with poor image quality were excluded from the analysis. We proposed a Scalable Multi Instance Transformer (SMIT), a deep learning model, to predict PD-L1 expression from chest computed tomography (CT) scans, thereby reducing the need for invasive biopsy procedures. Compared to prior studies, our approach integrates multi-scale features from CT images, enhancing prediction accuracy and robustness. SMIT achieved superior performance in predicting PD-L1 status with precision (0.82), sensitivity (0.83), F1 score (0.83), area under the curve (AUC; 81%), and Precision-Recall AUC (0.80). SMIT’s predictions for PD-L1 status (≥50% or < 50%) were comparable to those derived from IHC-based PD-L1 status, validating its potential as a non-invasive diagnostic tool. Additionally, SMIT’s predictions for progression-free survival (PFS) were on par with IHC-based predictions. The SMIT model represents a significant advancement in the non-invasive prediction of PD-L1 expression in NSCLC, offering a viable alternative to traditional biopsy methods. This innovation could streamline immunotherapy selection, making treatments more accessible and personalized.

Single-Source domain generalization in medical image segmentation has been studied as a more practical configuration to solve domain shift issues in clinical applications. Data augmentation plays an important role in improving the diversity of training data. Recent data augmentation methods aim to randomize or disrupt the texture of images to encourage models to focus more on shape features, which are considered domain-invariant. It’s worth noting that texture features such as intensity variations are crucial cues for distinguishing the boundaries between the tumor and normal tissues. However, these features are often disrupted or compromised in existing methods. To effectively leverage these texture features and enhance the performance of the model, we propose a novel anatomy-guided texture augmentation (AGTA) method. Specifically, as imaging parameters vary, different organs or tissues may exhibit varying changes in intensity, while the intensity variations within each organ or tissue tend to remain consistent. To simulate this, we partition different organs into distinct regions based on the anatomical information of the image. Each region is then assigned random variations. We evaluated our method against other SDG methods in cross-modality and cross-center cervical tumor segmentation experiments. Our results show that our method outperforms all competing methods by a large margin.

Immune checkpoint inhibitors (ICIs) are a cornerstone of modern oncological treatments, particularly in the management of various cancers through immunotherapy. Despite their clinical success, ICIs are often associated with several immune-related adverse events (irAEs), among which pneumonitis is particularly significant due to its potential severity. Accurately identifying patients at high-risk of developing ICI-induced pneumonitis remains a critical challenge in lung cancer patient management. Early detection and precise differentiation are essential for timely and appropriate therapeutic interventions, which can significantly alter patient outcomes. We developed the PANDA (Pneumonitis ANomaly Detection using AttentionU-Net) model to address this challenge, leveraging advanced deep learning techniques to improve the early predicting of ICI-induced pneumonitis. Baseline CT scans from 348 cases (33 pneumonitis cases) patients undergoing ICI therapy were analyzed to train and validate the model. The PANDA model utilizes the Attention U-Net architecture, incorporating attention mechanisms to enhance feature extraction and anomaly detection capabilities. Data augmentation techniques, including brightness normalization and pixel shuffling, were applied to improve model robustness. The model was trained on normal cases using an autoencoder-based method with anomaly detection through mean squared error (MSE) distribution, followed by testing on pneumonitis cases. The PANDA model demonstrated superior performance, achieving a precision of 0.76, sensitivity of 0.79, specificity of 0.79, F1-score of 0.78, AUC of 0.85 and a Precision-Recall AUC of 0.82. These results significantly outperform traditional models, including clinical and radiomics approaches. The clinical model, for instance, achieved a precision of 0.75, sensitivity of 0.67, specificity of 0.73, F1-score of 0.76, AUC of 0.69 and a precision-recall AUC of 0.76. The classical radiomics model showed improvements over the clinical model, with a precision of 0.81, sensitivity of 0.72, specificity of 0.80, F1-score of 0.76, AUC of 0.70 and a precision-recall AUC of 0.79, but still fell short of the PANDA model’s performance. These comparisons emphasize the enhanced predictive capacity of the deep learning approach, significantly outperforming traditional models.

Traditionally, identifying predictive biomarkers for treatment efficacy involves evaluating treatment-marker interactions through regression models considering clinical outcomes. This study seeks optimal personalized treatment strategies, known as individualized treatment rules (ITRs) through innovative approach utilizing weighting method and formulation of personalized treatment scoring. Integrating two scoring methods, (linear vs. non-linear) we aim to elucidate clinicogenomic indicators predictive of treatment efficacy. Emphasis is placed on identifying patients at elevated risk for early progression and determining their optimal treatment choice between immune checkpoint inhibitor-monotherapy (ICI-Mono) and ICI-Chemotherapy (ICI-Chemo). A total of 408 non-small cell lung cancer (NSCLC) patients, from MD Anderson and Mayo Clinic were enrolled. Performance was evaluated as average treatment effect of weighted risk reduction for 3-months progression between subgroup of patients who were treated according to vs. against model’s recommendation. The non-linear scoring method shows better performance comparing to the linear scoring method (overall risk reduction: −25.4% vs. −15.5% in training and −14.3 vs. −9.6% in testing cohort). Tobacco exposure and lung adenocarcinoma significantly influences outcomes in the ICI-Mono while stage-IVB and KRAS mutated gene associated with great effect from ICI-Chemo. These findings offer valuable insights for seamlessly integrating precision medicine into real-world clinical scenarios.

Dynamic Positron Emission Tomography (dPET) imaging and Time-Activity Curve (TAC) analyses are essential for understanding and quantifying the biodistribution of radiopharmaceuticals over time and space. Traditional compartmental modeling, while foundational, commonly struggles to fully capture the complexities of biological systems, including non-linear dynamics and variability. This study introduces an innovative data-driven neural network-based framework, inspired by Reaction Diffusion systems, designed to address these limitations. Our approach, which adaptively fits TACs from dPET, enables the direct calibration of diffusion coefficients and reaction terms from observed data, offering significant improvements in predictive accuracy and robustness over traditional methods, especially in complex biological scenarios. By more accurately modeling the spatio-temporal dynamics of radiopharmaceuticals, our method advances modeling of pharmacokinetic and pharmacodynamic processes, enabling new possibilities in quantitative nuclear medicine.

This study aims to evaluate the repeatability and reproducibility of radiomics features (RFs) under image perturbations and examine their generalizability across computed tomography (CT) and magnetic resonance (MR) images among nasopharyngeal carcinoma (NPC) patients. A total of 397 NPC patients with contrast-enhanced computed tomography (CECT), CET1-weight, and T2-weight MR images were analyzed. Image perturbation and contour randomization were implemented to the images and masks to mimic the scanning position and tumor segmentation stochasticity. A total of 1288 RFs from original, Laplacian-of-Gaussian-filtered (LoG) and wavelet-filtered images were extracted. The stability of RF was assessed by adopting median intraclass correlation coefficient (mICC) under patient subsampling. The mean absolute difference (MAD) of the mICC and the accuracy of the binarized repeatability between image datasets were adopted to evaluate its generalizability across image modalities. The MRI-based RFs showed higher stability (77.6% in CET1-w and 80.2% in T2-w with mICC ≥ 0.9), whereas the CT-based RFs were less stable (41.7% with mICC ≥ 0.9). Overall, 497 RFs (38.6%) had mICC ≥ 0.9 in all three modalities. Shape features consistently kept the highest stability in all modalities. MRI-based RFs displayed higher repeatability and reproducibility against scanning position and tumor segmentation variations than CT-based RFs. We urge caution when handling CT-based RFs and advice adopting MRI-based RFs with higher stability during feature pre-selection for stable model construction.