Open research questions in AI in cancer detection

What the literature leaves open

• Abstract Transfer learning offers a promising strategy for extending computational pathology models from data-rich to annotation-scarce cancer types, yet the transferability of tumor microenvironment (TME) representations across gastrointestinal organs remains underexplored.

  Cross-organ transfer learning for tumor microenvironment classification from colorectal to gastric cancer histopathology · 2026 · DOI

• Differentiating non-alcoholic steatohepatitis (NASH) from non-alcoholic fatty liver disease (NAFLD) using ultrasound remains challenging due to subtle tissue alterations and the limited information available in conventional B-mode imaging.

  Learning from Complementary Ultrasound Representations for Liver Disease Classification · 2026

• Although DL achieves high accuracy under noise-free conditions, real-world reliability remains underexplored because acquisition factors and speckle noise degrade images, increasing missed-malignancy risk.

  A Noise-Aware Robustness Evaluation Framework for Breast Cancer Classification in Ultrasound Imaging · 2026 · DOI

• One of the key limitations of this study is the exclusive use of the MIAS dataset for model training and evaluation. However, some research has found worse outcomes due to strong image noise, limited data or more simplistic designs.

  Early Breast Cancer Diagnosis in Mammography Through Microcalcification and Mass Segmentation with Deep Learning Classification · 2026 · DOI

• One key limitation is the retrospective design, which inherently carries the potential for selection bias. As such, the findings may be limited by the quality of the available data, and future prospective studies could help mitigate this issue. Without independent external or multi-institutional validation, the general- izability and robustness of the proposed models remain uncertain.

  Prediction of histological grading in ductal carcinoma in situ based on mammographic signs and clinical information using machine learning models · 2026 · DOI

• The advancement of AI has been significantly ac‐ celerated by the availability of large, publicly acces‐ sible datasets, which provide high‐quality training VOLUME 25, ISSUE 2, APRIL-JUNE 2026 Artificial Intelligence in Reproductive Medicine: Transforming Diagnosis, Treatment, and IVF Outcomes data for machine learning models. AI enables clini‐ cians to collect, process, and analyze complex datasets in previously unattainable ways, thereby en‐ hancing decision‐making in ART, including the selec‐ tion of optimal embryos and sperm. The integration of big data analytics offers the potential to derive ro‐ bust, evidence‐based insights from complex biologi‐ cal patterns. As these databases continue to expand and evolve, reproductive medicine is poised to expe‐ rience substantial progress. Comprehensive data analysis facilitates the extraction of actionable knowledge, and the performance of AI systems can be further augmented when combined with comple‐ mentary computational approaches, such as ma‐ chine learning, provided that high‐quality data collection, integration, and interpretation are en‐ sured. The future of AI in reproductive medicine de‐ pends on moving from conceptual promise to clinically testable hypotheses supported by rigorous validation. For embryo selection, CNN models trained on standardized, multi‐centre datasets may outperform embryologists in predicting implanta‐ tion potential, which can be tested through random‐ ized controlled trials comparing AI‐assisted versus conventional selection. In male infertility, hybrid AI models (e.g., SVM with deep learning) could achieve higher accuracy in assessing sperm motility and morphology, requiring prospective validation across diverse laboratories and demographics. For ovarian stimulation, reinforcement learning–based dosing al‐ gorithms are hypothesized to reduce OHSS incidence without compromising oocyte yield, a claim testable through clinical trials of AI‐guided versus physician‐ guided dosing. Idiopathic infertility may be approach by integrative big data combining genomics, imaging, and clinical records to identify novel biomarkers, which can be validated in longitudinal cohort studies with AI‐driven feature extraction. In reproductive surgery, AI‐augmented robotic systems are predicted to shorten operative time and reduce complication VOLUME 25, ISSUE 2, APRIL-JUNE 2026 rates, a hypothesis suitable for comparative effective‐ ness studies against current robotic techniques. Col‐ lectively, these testable directions underscore the importance of building large‐scale harmonized datasets, applying NLP and DL for knowledge extrac‐ tion, and embedding AI‐driven decision‐support sys‐ tems into clinical workflows in a safe, reproducible, and equitable manner.

  Artificial Intelligence in Reproductive Medicine: Transforming Diagnosis, Treatment, and IVF Outcomes · 2026 · DOI

• 4 Future work To further expand the clinical applicability of EGP-Net, future work will focus on the following key research directions. Future research will explore three-dimensional segmentation architectures to improve the model’s ability to characterize the three-dimensional morphology of nodules and their spatial interactions with adjacent tissues.

  EGP-Net: a lung nodule segmentation network integrating edge guidance and pyramidal multi-scale contextual attention mechanisms · 2026 · DOI

• Future work will focus on: (1) validating the models on larger, multi-institutional datasets to assess external generalizability; ENGINEERING MODELLING 39 (2026) 1, 1-18 15 Y. However, like many models evaluated on the WBCD dataset, their generalizability to external, multi-center clinical data remains to be validated.

  An Improved Early Breast Cancer Cells Classification and Prediction Based on a Fuzzy Neural Network Model · 2026 · DOI

• Because APP-positive profiles are sparse, fragmented, stain-variable, and morphologically diverse, PIGMENT combines limited expert labels with APP-specific augmentation designed to model variation in APP-positive intensity, size, continuity, fragmentation, and local tissue context.

  PIGMENT: A deep learning framework for Porcine Immunohistochemistry seGMENTation · 2026 · DOI

• Further prospective valida- tion using multi-institutional datasets and correlation with ISH/ FISH results is warranted to determine its clinical applicability in routine pathology workflows.

  Deep learning based automated HER2 score prediction using immunohistochemistry histopathological images: a dual-center study · 2026 · DOI

• Several limitations should be acknowledged. First, we did not directly compare our framework with task-specific models like HER2Net, as the pixel- and cell-level annotations they require were unavailable in our retrospective cohort. Second, our relatively small dual-center dataset requires larger external validation cohorts to confirm generalizability. Third, our patch-level approach pres- ents several constraints. Inheriting slide-level HER2 labels for patch-level training may introduce label noise due to intratumoral staining heterogeneity. Additionally, patch-level bootstrap resampling may slightly underestimate the true uncertainty of model performance due to intra-WSI patch correlations. Finally, without a slide- or patient-level aggregation strategy, our results represent preliminary patch-level evidence rather than clinical slide-level HER2 scoring. Future studies should utilize larger multi-center datasets, cell-level annotations, and comparisons with task-specific or hybrid weakly supervised models, while developing robust slide-level or patient-level aggregation method.

  Deep learning based automated HER2 score prediction using immunohistochemistry histopathological images: a dual-center study · 2026 · DOI

• In contrast, the application of AI in neurodegenerative and inflammatory conditions is constrained by the subtlety of the pathological features, the lack of standardized quantitative grading systems, and the scarcity of large, multi-institutional annotated datasets (Kim et al.

  Digital pathology and artificial intelligence in neuropathological diagnosis: a comprehensive review of current applications, challenges, and future directions · 2026 · DOI

• The data preprocessing stage involves three steps which include cleaning missing values and inconsistent data, normalizing attributes, and using SMOTE to balance class distribution in the dataset.

  Review on Machine Learning-Driven Cervical Cancer Prediction System · 2026 · DOI

• Extensive experiments on scar classification demonstrate that our method consistently outperforms end-to-end deep learning baselines or using LLMs as black-box classifiers under limited data conditions, establishing a promising direction for integrating LLMs into data-efficient and clinically transparent medical AI systems.

  When LLMs Analyze Scars: From Images to Clinically-Meaningful Features · 2026

• In this study, we present a leakage-safe stacking design that integrates (i) tabular features, (ii) KGE features, (iii) semantic subtyping signals, and (iv) SPARQL rule features, with reproducible templates and released code for breast cancer prediction and explanation. By leveraging domain-specific ontologies and SPARQL-based inference, we constructed enriched feature representations that capture both statistical and semantic relationships among clinical attributes. The proposed approach was validated on four benchmark datasets: WDBC, Coimbra, UMC Ljubljana, and METABRIC. Across all datasets, our model consistently outperformed conventional classifiers and recent related studies in terms of accuracy, precision, recall, F1-score, and ROC-AUC. Moreover, the use of PyKEEN for embedding generation and the application of Leiden and HDBSCAN clustering enabled transparent subgroup discovery and semantic explanation of patient clusters. These semantic clusters, further explained through SPARQL queries, enhance the interpretability and trustworthiness of the prediction results - an essential requirement in clinical decision-making environments. For future work, we plan to extend our framework to incorporating multi-modal biomedical data sources, imaging, genomic, and temporal data into the ontology schema. We also aim to investigate transformer-based embedding models such as OWL-BERT and Neuro- Symbolic Reasoning to further improve the generalization the predictions. Additionally, and explainability of integration with real-time clinical decision support systems will be explored to validate the practical deployment of our approach in hospital settings.

  Ontology-Enhanced Machine Learning Models for Breast Cancer Diagnosis · 2026 · DOI

• Although the proposed architecture achieves competitive performance relative to established baseline models while maintaining lower computational requirements, additional improvement remains necessary. The model has a classification accuracy that is 75.56%; this value denotes that additional optimization is necessary to use this model in high-stakes medical applications. Cur- rent experimental results are based on a training regime that excludes data balancing loss functions, sophisticated augmentation strategies, and extensive preprocessing, which likely limits the final predictive performance. Future research will be aimed at combining class-weighting methods and more accurate preprocessing procedures to enhance the overall accuracy and stability of the models. Moreover, even with the incorporated visual explanations of Gradient-weighted class activation mapping, the interpretability is limited due to the absence of expert-level pathological validation. Future studies need to focus on the implication of medical experts to compare the described discriminative nuclear areas with clinical biomarkers to guarantee biological importance and diagnostic value.

  A lightweight deep learning model with channel attention for kidney cell classification from microscopy images · 2026 · DOI

• Woloshuk et al. (2021) 21 Yang et al. (2023) 18 Doerrich et al. (2025) 20 Manzari et al. (2023) 23 Manzari et al. (2025) 24 Liu et al. (2024) 25 Bundele et al. (2025) 26 Sun et al. (2025) 27 Winfree et al. (2023) 29 Carnevali et al. (2024) 30 Mamalakis et al. (2024) 31 An et al. (2025) 32 Chanchal et al. (2025) 28 Ferdous et al.

  A lightweight deep learning model with channel attention for kidney cell classification from microscopy images · 2026 · DOI

• Looking back at the reviewed methods, the development trend of medical image-to-mesh reconstruction has shown a clear technological evolution path: from statistical shape models to template models, then to generative models, and finally to implicit models. Among these, implicit models have demonstrated significant advantages in both fidelity and regularization. Currently, the implicit models used for medical image reconstruction primarily include SDF and NeuralODE approaches. Meanwhile, in the fields of computer vision and com- puter graphics, several surface reconstruction algorithms that outperform SDF Yariv et al. [2021] have emerged. High-fidelity surface reconstruction using neural implicit reconstruction Wang et al. [2021, 2023a], Liu et al. [2023], Yariv et al. [2023] has become a hot research topic. Com- pared to SDF, Neural ODE and NeRF, new representation methods (such as Hash coding M¨uller et al. [2022], Li et al. [2023], Gaussian splatting Kerbl et al. [2023], Gao et al. [2025], Gu´edon and Lepetit [2024], Huang et al. [2024], Yu et al. [2024], Chen et al. [2024a], Dai et al. [2024], Wang et al. [2024], Tang et al. [2024]) provide better tex- ture preservation. Gaussian splatting technique has demonstrated break- through advantages in surface reconstruction. This ap- proach achieves superior computational efficiency with training time reduced by more than 10 times Lyu et al. [2024] compared to existing neural implicit reconstruction methods. Through explicit 3D Gaussian point representa- tion and surface alignment strategies, complex surface de- tails can be accurately reconstructed. The flexible optimiza- tion framework allows simultaneous optimization of geo- metric structure and appearance features, achieving high- fidelity reconstruction. Furthermore, its use of sparse 3D Gaussian points representation significantly reduces mem- ory overhead, while its end-to-end reconstruction capability avoids post-processing steps like depth fusion in traditional methods, thereby reducing error accumulation. These ad- vantages make Gaussian splatting show tremendous poten- tial in medical image reconstruction. In the future, combin- ing this efficient neural surface reconstruction technology with the specific requirements of medical image process- ing is expected to bring major breakthroughs in medical im- age reconstruction quality. Particularly in clinical applica- tion scenarios that require a balance of reconstruction accu- racy, computational efficiency, and real-time performance, the advantages of Gaussian splatting will become especially prominent.

  From pixels to polygons: A survey of deep learning approaches for medical image-to-mesh reconstruction · 2026 · DOI

• How our work addresses it 97.48% (CT, ROI 17×17) ROI-based; limited scalability for complex datasets End-to-end classification on full CT images, eliminating ROI dependency Hekmat (2) 98.7% Ensemble CNNs but no explainability, limiting clinical trust Integration of XAI (SHAP, Grad-CAM, Saliency) to provide interpretable outputs Naeem (6) 95.78% (MRI), 97.44% (CT), 99%…

  Leveraging deep learning and explainable AI for effective liver tumor classification from CT scan images · 2026 · DOI

• Another limitation of this study is the class imbalance of the dataset. Although current mainstream single deep learning-based aux- iliary diagnostic algorithms have demonstrated certain application value in medical image analysis, they are limited by the singularity of feature representation, resulting in obvious constraints on diag- nostic efficacy and stability in complex clinical scenarios.

  Hybrid deep feature and machine learning framework for classification of thyroid nodules in ultrasound images · 2026 · DOI

• observed differences were This study establishes a rigorous, leakage-free benchmarking for breast cancer histopathology classification, framework systematically evaluating nine state-of-the-art deep learning architectures. By enforcing a strict patient-aware cross-validation protocol, this work addresses a critical flaw in prior literature: the artificial inflation of model performance due to patient-level data leakage. Our comprehensive quantitative and statistical analyses reveal a fundamental insight: when evaluated under highly controlled, patient-isolated conditions, no statistically across modern significant Convolutional Neural Networks (e.g., ResNet50, DenseNet121) and Vision Transformers (e.g., Swin-Small). While DenseNet121 and ResNet50 demonstrated exceptional stability and high recall, achieving over 92% mean accuracy across 5-fold and 10-fold evaluations, hypothesis incremental architectural complexity did not yield statistically significant differences the magnification-wise analysis highlighted the critical role of spatial context. All models achieved their peak performance at lower magnifications is suffered a measurable decline at higher preserved, and), where limited fields of view obscure magnifications (400 contextual relationships. Ultimately, this study demonstrates that the integrity of the evaluation protocol and data partitioning strategy plays a more critical role in enabling reliable real-world deployment than the choice of baseline architecture. in diagnostic performance. Furthermore, tissue architecture testing confirmed), where global that (40 � � Based on these findings and the recognized limitations of single-scale patch classification, several promising avenues for future research emerge: Instance Learning (MIL) and multi-resolution fusion frameworks (such as Graph Neural Networks or state space models like SlideMamba). These architectures can simultaneously leverage fine-grained cellular atypia at high magnifications and global structural organization at low magnifications, mimicking the holistic approach of a human pathologist. • Cross-Domain Generalization and Multi-Centric Validation: While the models exhibited strong stability on the BreaKHis dataset, histopathological images are highly susceptible to domain shifts caused by variations in tissue preparation, staining protocols (H&E), and digital should prioritize external validation across diverse, multi-centric cohorts (e.g., TCGA or CAMELYON) and integrate domain adaptation or color normalization techniques to ensure robust real-world generalization. scanner calibrations.

  A patient-aware benchmarking of CNN and transformer architectures for breast cancer histopathology classification · 2026 · DOI

• learning. More recent architectures, such as SlideMamba [22], integrate graph-based representations with sequence modeling techniques to enhance contextual understanding across spatial scales. These approaches highlight a shift toward incorporating higher-order structural information, which is not explicitly modeled in standard CNN or patch-based transformer frameworks. to conventional [21] quantify feature Another emerging direction is the development of foundation models for computational pathology. These models leverage large- scale pretraining to learn transferable representations that can generalize across [23] demonstrates that such models can achieve strong performance across diverse histopathological benchmarks. However, their large-scale data and computational resources reliance on presents practical controlled experimental settings and smaller datasets such as BreaKHis. tasks and datasets. Recent work limitations, particularly for families. providing innovations across without different model Despite the diversity of approaches, several key limitations persist across the literature. First, many studies focus on specific controlled architectural comparisons Second, inconsistencies in preprocessing, training protocols, and evaluation strategies hinder reproducibility and fair benchmarking. Third, statistical validation of performance differences rarely conducted, limiting the ability to draw reliable conclusions regarding model superiority. Finally, while advanced paradigms such as MIL, spatial modeling, and foundation models offer improved representational capabilities, their complexity often restricts their use in standardized benchmarking scenarios. is In summary, while prior work has demonstrated the effectiveness of deep learning for histopathological classification, there remains a clear need for a systematic and controlled fair comparison across evaluation that enables framework for diverse architectures. Such a is essential framework understanding whether observed performance differences are attributable to model design or experimental variability.

  A patient-aware benchmarking of CNN and transformer architectures for breast cancer histopathology classification · 2026 · DOI

• The synergy of mammogram and ultrasound supported with AI-reading mammogram may provide the best performance especially when careful clinical correla- tion is applied to interpret borderline findings.

  Operational insights into AI-assisted mammographic interpretation for enhanced sono-mammography performance · 2026 · DOI

• Despite substantial progress, several challenges continue to limit the widespread clinical implementation of AI in breast pathology. These include variability in data quality and an- notation standards, limited dataset diversity, potential algo- rithmic bias, and concerns regarding model generalizability across institutions and populations. Additional barriers in- clude infrastructure and computational requirements, inte- gration into existing pathology workflows, and the need to maintain patient privacy, data security, and ethical stand- ards. Furthermore, issues related to interpretability, regula- tory approval, reimbursement, and standardized validation protocols remain critical considerations for safe and equitable adoption.

  Artificial Intelligence in Breast Pathology: Recent Advances in Multimodal Models, Explainability, and Clinical Applications · 2026 · DOI

• Future efforts will focus on developing interpretable, clinically validated, and seamlessly integrated AI systems that complement pathologists in routine breast pathology practice. Emerging technologies such as multimodal foundation models, generative AI, and spatial TME analysis are expected to enable more robust prediction of treatment response, recurrence risk, and patient outcomes through integration Journal of Clinical and Translational Pathology 2026 7 Table 1.

  Artificial Intelligence in Breast Pathology: Recent Advances in Multimodal Models, Explainability, and Clinical Applications · 2026 · DOI

Most-cited papers in AI in cancer detection

• Segment anything in medical images · Nature Communications · 2024 · 2,440 citations
• Towards a general-purpose foundation model for computational pathology · Nature Medicine · 2024 · 1,185 citations
• A whole-slide foundation model for digital pathology from real-world data · Nature · 2024 · 656 citations
• ASF-YOLO: A novel YOLO model with attentional scale sequence fusion for cell instance segmentation · Image and Vision Computing · 2024 · 446 citations
• A pathology foundation model for cancer diagnosis and prognosis prediction · Nature · 2024 · 443 citations
• A foundation model for clinical-grade computational pathology and rare cancers detection · Nature Medicine · 2024 · 400 citations
• A multimodal generative AI copilot for human pathology · Nature · 2024 · 367 citations
• CellViT: Vision Transformers for precise cell segmentation and classification · Medical Image Analysis · 2024 · 287 citations
• A Multimodal Biomedical Foundation Model Trained from Fifteen Million Image–Text Pairs · NEJM AI · 2024 · 206 citations
• A novel lightweight deep convolutional neural network for early detection of oral cancer · Oral Diseases · 2021 · 168 citations

Most recent work

• Commercially Available Artificial Intelligence Score on Preoperative Mammography for Prediction of Future Breast Cancer After DCIS Treatment · American Journal of Roentgenology · 2026
• Mammography-based artificial intelligence model for predicting axillary lymph node status after neoadjuvant therapy in breast cancer · European Radiology · 2026
• An integrated automated deep learning framework for annotating tumor-infiltrating lymphocytes in lung adenocarcinoma pathology · Frontiers in Bioinformatics · 2026
• Real-World Single-Reading Screening Mammography Performance When Using an FDA-Approved Artificial Intelligence Tool · American Journal of Roentgenology · 2026
• Transformers meet CNNs for insights into breast mass classification from histopathological images · Frontiers in Artificial Intelligence · 2026
• Feature-optimized deep neural network with explainable AI for lung cancer histopathology classification · Discover Artificial Intelligence · 2026
• Artificial intelligence reshaping the paradigm of hematologic malignancy diagnosis and treatment: From static assessment to dynamic precision management · Annals of Hematology · 2026
• Subspecialty-specific foundation model for intelligent gastrointestinal pathology · npj Digital Medicine · 2026
• Transfer Learning in Convolutional Neural Network to Differentiate Follicular Adenoma Versus Follicular Carcinoma of Thyroid on Aspiration Cytology Material · Cytopathology · 2026
• SortIT - A Tool For Assessing Observer Variability And Creating Ground Truth Image Classification Datasets · bioRxiv · 2026

Related topics in Computer Science