https://thisweekinmedicalimagingai.com/The highest-impact medical imaging AI papers, re-ranked weekly, with machine-written summaries checked against their source.enhttps://thisweekinmedicalimagingai.com/papers/W2763355946https://thisweekinmedicalimagingai.com/papers/W2763355946Nature Reviews Clinical Oncology2017-10-04**Standardizing Radiomics for Clinical Utility** This review identifies the specific technical and reporting hurdles preventing the transition of radiomic features—extracted from standard-of-care imaging—into reliable clinical-decision support systems. For the practitioner or researcher, it provides a framework for moving beyond 'big data' mining toward a disciplined methodology for predicting patient outcomes and treatment responses in oncology. It specifically addresses the need for standardized evaluation criteria to ensure that high-throughput image features translate into reproducible, chttps://thisweekinmedicalimagingai.com/papers/W2803760365https://thisweekinmedicalimagingai.com/papers/W2803760365Nature reviews. Cancer2018-05-17**Establishing the Framework for AI in Oncology Imaging** This review synthesizes how deep learning architectures, such as convolutional neural networks and variational autoencoders, shift the paradigm of radiological analysis from subjective, qualitative visual assessment to objective, quantitative reporting. While not a primary research study, it provides a foundational overview of how these automated patterns can be integrated into clinical oncology workflows, specifically regarding disease detection, characterization, and monitoring.https://thisweekinmedicalimagingai.com/papers/W2174661749https://thisweekinmedicalimagingai.com/papers/W2174661749Radiology2015-11-18**Defining the Transition from Visual Interpretation to Data Mining in Oncology** This overview establishes the conceptual framework for radiomics by defining the systematic extraction of high-order features—beyond what is visible to the human eye—from standard-of-care medical images. For the clinic, the shift involves moving from qualitative visual interpretation to the mining of quantitative data to build predictive models for cancer diagnosis and prognosis. The text highlights that because these features are derived from routine imaging, the workflow aims to integrate automated data analyshttps://thisweekinmedicalimagingai.com/papers/W2767128594https://thisweekinmedicalimagingai.com/papers/W2767128594Cancer Research2017-10-31**Standardizing Radiomic Feature Extraction with PyRadiomics** For researchers and physicists seeking to move beyond custom, non-reproducible scripts, this paper introduces **PyRadiomics**, a Python-based open-source platform designed to standardize the extraction of high-dimensional features from medical images. While the paper focuses on lung lesion characterization, the utility lies in its architecture: it provides a consistent framework for engineered features (radiomics) to overcome the reproducibility hurdles currently hindering the validation of noninvasive imaging biomarkers. The toolhttps://thisweekinmedicalimagingai.com/papers/W2998789541https://thisweekinmedicalimagingai.com/papers/W2998789541Radiology2020-03-10**Standardizing Radiomics for High-Throughput Phenotyping** The Image Biomarker Standardization Initiative (IBSI) provides a framework for standardizing the extraction and reporting of radiomic features across various imaging modalities and tasks. Instead of inconsistent reporting, the initiative establishes standardized terminology and workflows for feature calculation, ensuring that quantitative metrics remains reproducible across different software platforms and clinical sites. This is a critical step for researchers and clinicians seeking to integrate high-throughput radiomics into clinichttps://thisweekinmedicalimagingai.com/papers/W4212875960https://thisweekinmedicalimagingai.com/papers/W42128759602022 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)2022-01-01**UNETR: Integrating Transformers into the U-Net Architecture for 3D Segmentation** This architecture addresses the limitation of local receptive fields in standard convolutional neural networks (FCNNs) by incorporating a transformer-based encoder to capture long-range spatial dependencies in volumetric data. By reframing 3D medical image segmentation as a sequence-to-sequence prediction task, the authors integrate the global scope of transformers while maintaining the hierarchical feature extraction of the U-shaped network design. The model achieved state-of-the-art performance on the BTCV mhttps://thisweekinmedicalimagingai.com/papers/W2751069891https://thisweekinmedicalimagingai.com/papers/W2751069891Scientific Data2017-09-05**Expert-Verified Segmentation Labels and Radiomics for TCGA Glioma MRI Collections** This dataset provides a standardized foundation for quantitative imaging research on 243 pre-operative multimodal MRI scans of glioblastomas and low-grade gliomas. The authors combined automated segmentation with manual refinement by a board-certified neuroradiologist to produce 'gold standard' labels for tumor sub-regions. For researchers, the release includes a pre-calculated, extensive panel of radiomic features mapped to these expert-validated segments, facilitating reproducible radiogenomic studies and https://thisweekinmedicalimagingai.com/papers/W2900954917https://thisweekinmedicalimagingai.com/papers/W2900954917Zeitschrift für Medizinische Physik2018-12-13**Mapping the Deep Learning Landscape in MRI** This review addresses the explosion of deep neural network applications within the MRI pipeline, providing a structured roadmap for those looking to move from theoretical understanding to practical application. Rather than focusing on a single algorithm, it covers the entire processing chain—including acquisition, image retrieval, segmentation, and disease prediction—while identifying specific open-source tools and datasets for researchers looking to begin implementing these models in clinical workflows.https://thisweekinmedicalimagingai.com/papers/W2911605224https://thisweekinmedicalimagingai.com/papers/W2911605224CA A Cancer Journal for Clinicians2019-02-05**Mapping the AI Landscape in Oncology Imaging** This review synthesizes the current state of AI applications across four specific cancer types: lung, brain, breast, and prostate. For clinicians and researchers, it identifies how AI can move beyond simple image processing to address complex tasks such as volumetric tumor delineation over time, predicting clinical outcomes from radiographic phenotypes, and assessing treatment impacts on adjacent organs. While the authors note that many existing studies lack rigorous validation for reproducibility, the review categorizes how AI is being positiohttps://thisweekinmedicalimagingai.com/papers/W3048802680https://thisweekinmedicalimagingai.com/papers/W3048802680Insights into Imaging2020-08-12**A 'How-to' Guide for Navigating Radiomics Workflows** This review addresses the technical hurdles and standardized workflows required to move radiomics from theoretical research to reproducible clinical application. For the reader, it provides a practical roadmap for implementing quantitative texture analysis and a critical discussion on how technical variability affects the reliability of extracted features across different imaging modalities.https://thisweekinmedicalimagingai.com/papers/W3172681723https://thisweekinmedicalimagingai.com/papers/W3172681723Nature Communications2022-07-15**Benchmarking Cross-Task Generalizability in Medical Image Segmentation** While most medical imaging research focuses on optimizing for a specific clinical niche, this study evaluates whether a single model architecture can perform reliably across diverse modalities and tasks. By evaluating a broad suite of datasets—including CT, MRI, and ultrasound—the researchers found that models achieving consistent performance across multiple distinct tasks (e.g., organ segmentation in CT and vessel segmentation in MRI) demonstrate superior generalization to unseen clinical problems compared to models thttps://thisweekinmedicalimagingai.com/papers/W1408981388https://thisweekinmedicalimagingai.com/papers/W1408981388Scientific Reports2015-08-17**Evaluating Stability and Performance of Machine Learning for Lung Cancer Survival Prediction} 14 feature selection methods and 12 classification models were tested to identify the most robust ways to extract predictive biomarkers from pre-treatment CT scans. The study involved 440 radiomic features across a cohort of 464 patients to determine which specific combinations provide the most stable results against data perturbation. The authors found that the Wilcoxon-based selection method (WLCX) and the random forest (RF) classifier provided the highest prognostic performance for predicting ovehttps://thisweekinmedicalimagingai.com/papers/W2611467245https://thisweekinmedicalimagingai.com/papers/W2611467245Nature2018-03-01**Neural Networks as a Unified Framework for Multi-Modal Image Reconstruction** Current reconstruction pipelines for MRI, CT, and PET often rely on handcrafted signal processing chains that require extensive manual tuning to account for noise and sensor non-idealities. This paper introduces AUTOMAP, a deep learning framework that replaces these multi-stage pipelines with a single, data-driven mapping between the sensor domain and the image domain. The authors demonstrate that a single network architecture can handle diverse MRI acquisition strategies while remaining robust to noise and reducihttps://thisweekinmedicalimagingai.com/papers/W2913559493https://thisweekinmedicalimagingai.com/papers/W2913559493Engineering2019-01-29**Deep Learning in Medical Ultrasound Analysis** This review synthesizes the application of deep learning architectures specifically for ultrasound image analysis. Given the inherent challenges of US imaging—such as low image quality and high operator variability—the authors categorize current deep learning methods into three primary clinical tasks: classification, detection, and segmentation. It serves as a roadmap for understanding how neural networks are being deployed to increase the objectivity and accuracy of ultrasound-based diagnoses.https://thisweekinmedicalimagingai.com/papers/W2751538714https://thisweekinmedicalimagingai.com/papers/W2751538714Scientific Reports2017-08-29**Deep Learning vs. Handcrafted Radiomics for GBM Prognosis** This study evaluates whether deep learning-based features can outperform traditional, hand-engineered radiomics for predicting overall survival (OS) in Glioblastoma Multiforme (GBM). Using preoperative multi-modality MRI, the authors compared 98,304 deep features (via transfer learning) against 1,403 handcrafted features. The resulting six-feature deep learning signature outperformed traditional clinical risk factors alone (C-index 0.710 vs. baseline) and reached a C-index of 0.739 when combined with clinical variables like age andhttps://thisweekinmedicalimagingai.com/papers/W4223430324https://thisweekinmedicalimagingai.com/papers/W4223430324npj Digital Medicine2022-04-12**Addressing Systematic Biases in Medical Imaging AI** While machine learning in medical imaging shows promise, this review identifies the systematic methodological failures—ranging from data bias to publishing incentives—that currently hinder clinical translation. It provides a framework for identifying these pitfalls and outlines specific recommendations for more robust study design and evaluation.https://thisweekinmedicalimagingai.com/papers/W2193325189https://thisweekinmedicalimagingai.com/papers/W2193325189Procedia Computer Science2015-01-01**Comparative Analysis of Block-Based Segmentation Techniques** This survey categorizes and compares block-based segmentation methods, which are foundational for defining specific regions of interest within medical imagery. While the paper does not focus on a single clinical modality, it evaluates the integration of domain knowledge—a critical factor for radiologists and researchers moving from general algorithmic outputs to clinically relevant automated segmentations.https://thisweekinmedicalimagingai.com/papers/W2600642189https://thisweekinmedicalimagingai.com/papers/W2600642189Scientific Reports2017-08-25**Integrating Radiomics with Clinical Data to Predict Treatment Failure in H&N Cancer** This study evaluates whether 1,615 radiomic features extracted from pre-treatment FDG-PET and CT scans can improve the prediction of locoregional recurrence (LR) and distant metastasis (DM) in head-and-neck cancer. By combining these high-dimensional image features with clinical variables using random forest models, the researchers achieved an AUC of 0.69 for predicting LR and a significantly higher AUC of 0.86 for predicting DM. The results suggest that radiomics may provide a quantifiable way to identifyhttps://thisweekinmedicalimagingai.com/papers/W4401725766https://thisweekinmedicalimagingai.com/papers/W4401725766IEEE Transactions on Pattern Analysis and Machine Intelligence2024-08-21**Navigating the U-Net Landscape: A Taxonomy and Implementation Library** For researchers looking to move beyond standard U-Net architectures, this review provides a structured taxonomy of model variants—including Transformer-hybrid architectures and probabilistic prediction methods—across diverse medical imaging modalities. To assist in model selection, the authors include a curated list of clinical evaluations on well-known datasets and a comprehensive implementation library to streamline the transition from research to practical deployment.https://thisweekinmedicalimagingai.com/papers/W1909740415https://thisweekinmedicalimagingai.com/papers/W1909740415BMC Medical Imaging2015-08-11**Standardizing 3D Segmentation Metrics and Evaluation Tools** Selecting the correct metric to evaluate 3D segmentation performance is often complicated by inconsistent definitions in literature and the high computational cost of processing large volumes like whole-body MRI or CT. This work addresses these issues by providing a guide for selecting appropriate metrics and releasing an open-source tool that implements 20 different metrics—including specific versions for fuzzy segmentation—optimized for speed and memory efficiency on large datasets.https://thisweekinmedicalimagingai.com/papers/W3205076722https://thisweekinmedicalimagingai.com/papers/W3205076722Nature Reviews Clinical Oncology2021-10-18**Assessing the Clinical Utility of Radiomics and Deep Learning in Oncology}** This perspective outlines how AI-integrated imaging analysis can move beyond simple diagnosis to address complex clinical decision-making in oncology. The authors categorize specific high-value applications, including predicting treatment response, identifying true progression from benign treatment effects, and forecasting mutational profiles from imaging data. For the clinical team, the review provides a framework for evaluating the distinctions between handcrafted radiomic features and deep learning-based represehttps://thisweekinmedicalimagingai.com/papers/W2587297900https://thisweekinmedicalimagingai.com/papers/W2587297900Annals of Oncology2017-02-07**Navigating the path to clinical integration for radiomics** This review addresses the transition of radiomics from a research methodology to a practical clinical decision-support tool in oncology. While the authors acknowledge the potential of extracting quantitative markers to characterize tumor biology and spatial heterogeneity, they outline the specific standardization and refinement hurdles that must be overcome before these biomarkers can be integrated into routine clinical workflows. This is relevant for departments evaluating how automated image processing can contribute to precisionhttps://thisweekinmedicalimagingai.com/papers/W2327203407https://thisweekinmedicalimagingai.com/papers/W2327203407Scientific Reports2016-03-24**Assessing the Impact of CT Reconstruction Protocols on Radiomic Stability** This study addresses a critical hurdle in the clinical adoption of radiomics: the sensitivity of quantitative features to technical acquisition variables. By using a same-day repeat scan dataset of lung cancer patients, the authors test how variations in slice thickness (1.25 mm, 2.5 mm, and 5 mm) and reconstruction algorithms (sharp vs. smooth) influence the stability of radiomic features. While the results indicate that many features remain reproducible across a variety of settings, the study specifically identifihttps://thisweekinmedicalimagingai.com/papers/W2898197178https://thisweekinmedicalimagingai.com/papers/W2898197178Medical Physics2018-10-27**Survey of Deep Learning Methodologies in Radiation Therapy** This review outlines the transition of deep learning from a theoretical framework to a practical tool in clinical workflows. For practitioners and researchers, it provides a structured overview of five major application areas in medical imaging and radiation therapy, while specifically detailing methods for expanding small medical datasets—a common bottleneck in moving models from sandbox to clinic.https://thisweekinmedicalimagingai.com/papers/W2789588857https://thisweekinmedicalimagingai.com/papers/W2789588857NeuroImage2018-03-20**3D c-GANs for Low-Dose PET Reconstruction** This study addresses the trade-off between radiation safety and image quality in PET imaging. The authors propose a 3D conditional generative adversarial network (3D c-GAN) to synthesize high-quality, full-dose PET images from low-dose inputs. By incorporating a 3D U-Net architecture with skip connections and a progressive refinement scheme, the model aims to minimize noise and reconstruct anatomical detail. Evaluation on a human brain dataset involving both healthy subjects and those with mild cognitive impairment (MCI) indicates that the 3D c-GAhttps://thisweekinmedicalimagingai.com/papers/W2891590469https://thisweekinmedicalimagingai.com/papers/W2891590469Medical Image Analysis2018-12-08**Unsupervised Deep Learning for Faster MRI and CT Alignment** Manual or intensity-based image registration is computationally intensive, especially for large deformation tasks. This study introduces the DLIR framework, which removes the requirement for pre-labeled ground-truth data to train convolutional neural networks (ConvNets). Because the model learns from image similarity rather than manually annotated pairs, it can be deployed for one-shot registration of unseen images.https://thisweekinmedicalimagingai.com/papers/W3135096391https://thisweekinmedicalimagingai.com/papers/W3135096391Physica Medica2021-03-01**Navigating the practical hurdles of clinical AI integration** This review outlines the technical requirements and methodological pitfalls involved in transitioning AI from research prototypes to clinical decision support systems. While many papers focus on specific algorithms, this review synthesizes the infrastructure required for reliable deployment, including data harmonization to mitigate protocol-induced noise, the use of federated learning for multi-center data, and methods for addressing the 'black box' interpretability issue in deep learning models. It provides a useful framework fohttps://thisweekinmedicalimagingai.com/papers/W2785645041https://thisweekinmedicalimagingai.com/papers/W2785645041Journal of the American College of Radiology2018-02-04**Defining Success Metrics for AI Integration in Radiology** This review addresses the practical integration of AI into clinical workflows rather than just the underlying algorithms. For the practitioner, it outlines three specific use cases: triaging worklists to prioritize suspicious cases, extracting 'radiomic' features beyond human visual perception to improve prognosis, and automating routine tasks to improve work-life quality. The authors argue that the primary hurdle is not the replacement of radiologists, but the need for standardized nomenclature, better data-sharing methods, and righttps://thisweekinmedicalimagingai.com/papers/W2253150690https://thisweekinmedicalimagingai.com/papers/W2253150690Scientific Reports2015-06-05**Identifying Cancer-Specific Radiomic Clusters in Lung and Head & Neck CTs** This study evaluates how 3D texture and shape features from pre-treatment CT scans can be clustered to identify distinct tumor phenotypes in lung and head and neck cancers. By analyzing 878 patients, the researchers identified stable, cancer-specific radiomic clusters that correlate with clinical parameters including tumor stage, histology, and HPV status. The use of consensus clustering demonstrated high reproducibility (Rand Score of 0.92) across both cohorts, suggesting that specific sub-groups of radiomic featurhttps://thisweekinmedicalimagingai.com/papers/W3205594709https://thisweekinmedicalimagingai.com/papers/W3205594709Nature reviews. Cancer2021-10-18**Multimodal AI Models for Precision Oncology** Integrating radiological imaging, histology, and clinical data into a single predictive framework offers a path toward more nuanced biomarkers than single-modality analysis. This review outlines the specific engineering and computational hurdles—such as data sparsity and heterogeneity—that must be overcome to move beyond existing genomics standards and into integrated multimodal machine learning.