Medical image classification is crucial for diagnosis and treatment, benefiting significantly from advancements in artificial intelligence. The paper reviews recent progress in the field, focusing on three levels of solutions: basic, specific, and applied. It highlights advances in traditional methods using deep learning models like Convolutional Neural Networks and Vision Transformers, as well as state-of-the-art approaches with Vision Language Models. These models tackle the issue of limited labeled data, and enhance and explain predictive results through Explainable Artificial Intelligence.

Deep neural networks for medical image segmentation are often overconfident, compromising both reliability and clinical utility. In this work, we propose differentiable formulations of marginal L1 Average Calibration Error (mL1-ACE) as an auxiliary loss that can be computed on a per-image basis. We compare both hard- and soft-binning approaches to directly improve pixel-wise calibration. Our experiments on four datasets (ACDC, AMOS, KiTS, BraTS) demonstrate that incorporating mL1-ACE significantly reduces calibration errors, particularly Average Calibration Error (ACE) and Maximum Calibration Error (MCE), while largely maintaining high Dice Similarity Coefficients (DSCs). We find that the soft-binned variant yields the greatest improvements in calibration, over the Dice plus cross-entropy loss baseline, but often compromises segmentation performance, with hard-binned mL1-ACE maintaining segmentation performance, albeit with weaker calibration improvement. To gain further insight into calibration performance and its variability across an imaging dataset, we introduce dataset reliability histograms, an aggregation of per-image reliability diagrams. The resulting analysis highlights improved alignment between predicted confidences and true accuracies. Overall, our approach not only enhances the trustworthiness of segmentation predictions but also shows potential for safer integration of deep learning methods into clinical workflows. We share our code here: https://github.com/cai4cai/Average-Calibration-Losses

Recent work has shown improved lesion detectability and flexibility to reconstruction hyperparameters (e.g. scanner geometry or dose level) when PET images are reconstructed by leveraging pre-trained diffusion models. Such methods train a diffusion model (without sinogram data) on high-quality, but still noisy, PET images. In this work, we propose a simple method for generating subject-specific PET images from a dataset of multi-subject PET-MR scans, synthesizing "pseudo-PET" images by transforming between different patients' anatomy using image registration. The images we synthesize retain information from the subject's MR scan, leading to higher resolution and the retention of anatomical features compared to the original set of PET images. With simulated and real [$^{18}$F]FDG datasets, we show that pre-training a personalized diffusion model with subject-specific "pseudo-PET" images improves reconstruction accuracy with low-count data. In particular, the method shows promise in combining information from a guidance MR scan without overly imposing anatomical features, demonstrating an improved trade-off between reconstructing PET-unique image features versus features present in both PET and MR. We believe this approach for generating and utilizing synthetic data has further applications to medical imaging tasks, particularly because patient-specific PET images can be generated without resorting to generative deep learning or large training datasets.

The development of therapeutics builds on testing their efficiency in vitro. To optimize gene therapies, for example, fluorescent reporters expressed by treated cells are typically utilized as readouts. Traditionally, their global fluorescence signal has been used as an estimate of transduction efficiency. However, analysis in individual cells within a living 3D tissue remains a challenge. Readout on a single-cell level can be realized via fluo-rescence-based flow cytometry at the cost of tissue dissociation and loss of spatial information. Complementary, spatial information is accessible via immunofluorescence of fixed samples. Both approaches impede time-dependent studies on the delivery of the vector to the cells. Here, quantitative 3D characterization of viral transduction efficiencies in living retinal organoids is introduced. The approach combines quantified gene delivery efficiency in space and time, leveraging human retinal organ-oids, engineered adeno-associated virus (AAV) vectors, confocal live imaging, and deep learning-based image segmentation. The integration of these tools in an organoid imaging and analysis pipeline allows quantitative testing of future treatments and other gene delivery methods. It has the potential to guide the development of therapies in biomedical applications.

Adjuvant chemotherapy provides a limited survival benefit (<5%) for patients with stage II colorectal cancer (CRC) and is suggested for high-risk patients. Given the heterogeneity of stage II CRC, we aimed to develop a clinically explainable artificial intelligence (AI)-powered analyser to identify radiological phenotypes that would benefit from chemotherapy. Multimodal data from patients with CRC across six cohorts were collected, including 405 patients from the Guangdong Provincial People's Hospital for model development and 153 patients from the Yunnan Provincial Cancer Centre for validation. RNA sequencing data were used to identify the differentially expressed genes in the two radiological clusters. Histopathological patterns were evaluated to bridge the gap between the imaging and genetic information. Finally, we investigated the discovered morphological patterns of mouse models to observe imaging features. The survival benefit of chemotherapy varied significantly among the AI-powered radiological clusters [interaction hazard ratio (iHR) = 5.35, (95% CI: 1.98, 14.41), adjusted P_interaction = 0.012]. Distinct biological pathways related to immune and stromal cell abundance were observed between the clusters. The observation only (OO)-preferable cluster exhibited higher necrosis, haemorrhage, and tortuous vessels, whereas the adjuvant chemotherapy (AC)-preferable cluster exhibited vessels with greater pericyte coverage, allowing for a more enriched infiltration of B, CD4⁺-T, and CD8⁺-T cells into the core tumoural areas. Further experiments confirmed that changes in vessel morphology led to alterations in predictive imaging features. The developed explainable AI-powered analyser effectively identified patients with stage II CRC with improved overall survival after receiving adjuvant chemotherapy, thereby contributing to the advancement of precision oncology. This work was funded by the National Science Fund of China (81925023, 82302299, and U22A2034), Guangdong Provincial Key Laboratory of Artificial Intelligence in Medical Image Analysis and Application (2022B1212010011), and High-level Hospital Construction Project (DFJHBF202105 and YKY-KF202204).

Ultrasound-based bone surface segmentation is crucial in computer-assisted orthopedic surgery. However, ultrasound images have limitations, including a low signal-to-noise ratio, acoustic shadowing, and speckle noise, which make interpretation difficult. Existing deep learning models for bone segmentation rely primarily on costly manual labeling by experts, limiting dataset size and model generalizability. Additionally, the complexity of ultrasound physics and acoustic shadow makes the images difficult for humans to interpret, leading to incomplete labels in low-intensity and anechoic regions and limiting model performance. To advance the state-of-the-art in ultrasound bone segmentation and establish effective model benchmarks, larger and higher-quality datasets are needed. We propose a methodology for collecting ex-vivo ultrasound datasets with automatically generated bone labels, including anechoic regions. The proposed labels are derived by accurately superimposing tracked bone Computed Tomography (CT) models onto the tracked ultrasound images. These initial labels are refined to account for ultrasound physics. To clinically evaluate the proposed method, an expert physician from our university hospital specialized in orthopedic sonography assessed the quality of the generated bone labels. A neural network for bone segmentation is trained on the collected dataset and its predictions are compared to expert manual labels, evaluating accuracy, completeness, and F1-score. We collected UltraBones100k, the largest known dataset comprising 100k ex-vivo ultrasound images of human lower limbs with bone annotations, specifically targeting the fibula, tibia, and foot bones. A Wilcoxon signed-rank test with Bonferroni correction confirmed that the bone alignment after our optimization pipeline significantly improved the quality of bone labeling (p<0.001). The model trained on UltraBones100k consistently outperforms manual labeling in all metrics, particularly in low-intensity regions (at a distance threshold of 0.5 mm: 320% improvement in completeness, 27.4% improvement in accuracy, and 197% improvement in F1 score) CONCLUSION:: This work is promising to facilitate research and clinical translation of ultrasound imaging in computer-assisted interventions, particularly for applications such as 2D bone segmentation, 3D bone surface reconstruction, and multi-modality bone registration.

This study aimed to develop and validate machine learning models for preoperative identification of metastasis to station 4 mediastinal lymph nodes (MLNM) in non-small cell lung cancer (NSCLC) patients at pathological N0-N2 (pN0-pN2) stage, thereby enhancing the precision of clinical decision-making. We included a total of 356 NSCLC patients at pN0-pN2 stage, divided into training (n = 207), internal test (n = 90), and independent test (n = 59) sets. Station 4 mediastinal lymph nodes (LNs) regions of interest (ROIs) were semi-automatically segmented on venous-phase computed tomography (CT) images for radiomics feature extraction. Using least absolute shrinkage and selection operator (LASSO) regression to select features with non-zero coefficients. Four machine learning algorithms-decision tree (DT), logistic regression (LR), random forest (RF), and support vector machine (SVM)-were employed to construct radiomics models. Clinical predictors were identified through univariate and multivariate logistic regression, which were subsequently integrated with radiomics features to develop combined models. Models performance were evaluated using receiver operating characteristic (ROC) analysis, calibration curves, decision curve analysis (DCA), and DeLong's test. Out of 1721 radiomics features, eight radiomics features were selected using LASSO regression. The RF-based combined model exhibited the strongest discriminative power, with an area under the curve (AUC) of 0.934 for the training set and 0.889 for the internal test set. The calibration curve and DCA further indicated the superior performance of the combined model based on RF. The independent test set further verified the model's robustness. The combined model based on RF, integrating radiomics and clinical features, effectively and non-invasively identifies metastasis to the station 4 mediastinal LNs in NSCLC patients at pN0-pN2 stage. This model serves as an effective auxiliary tool for clinical decision-making and has the potential to optimize treatment strategies and improve prognostic assessment for pN0-pN2 patients. Not applicable.

Cone-beam CT (CBCT) scans that are affected by motion often require motion compensation to reduce artifacts or to reconstruct 4D (3D+time) representations of the patient. To do so, most existing strategies rely on some sort of gating strategy that sorts the acquired projections into motion bins. Subsequently, these bins can be reconstructed individually before further post-processing may be applied to improve image quality. While this concept is useful for periodic motion patterns, it fails in case of non-periodic motion as observed, for example, in irregularly breathing patients. To address this issue and to increase temporal resolution, we propose the deep single angle-based motion compensation (SAMoCo). To avoid gating, and therefore its downsides, the deep SAMoCo trains a U-net-like network to predict displacement vector fields (DVFs) representing the motion that occurred between any two given time points of the scan. To do so, 4D clinical CT scans are used to simulate 4D CBCT scans as well as the corresponding ground truth DVFs that map between the different motion states of the scan. The network is then trained to predict these DVFs as a function of the respective projection views and an initial 3D reconstruction. Once the network is trained, an arbitrary motion state corresponding to a certain projection view of the scan can be recovered by estimating DVFs from any other state or view and by considering them during reconstruction. Applied to 4D CBCT simulations of breathing patients, the deep SAMoCo provides high-quality reconstructions for periodic and non-periodic motion. Here, the deviations with respect to the ground truth are less than 27 HU on average, while respiratory motion, or the diaphragm position, can be resolved with an accuracy of about 0.75 mm. Similar results were obtained for real measurements where a high correlation with external motion monitoring signals could be observed, even in patients with highly irregular respiration. The ability to estimate DVFs as a function of two arbitrary projection views and an initial 3D reconstruction makes deep SAMoCo applicable to arbitrary motion patterns with single-view temporal resolution. Therefore, the deep SAMoCo is particularly useful for cases with unsteady breathing, compensation of residual motion during a breath-hold scan, or scans with fast gantry rotation times in which the data acquisition only covers a very limited number of breathing cycles. Furthermore, not requiring gating signals may simplify the clinical workflow and reduces the time needed for patient preparation.

Prenatal healthcare development requires accurate automated techniques for fetal ultrasound image segmentation. This approach allows standardized evaluation of fetal development by minimizing time-exhaustive processes that perform poorly due to human intervention. This research develops a segmentation framework through U-Net++ with ResNet backbone features which incorporates attention components for enhancing extraction of features in low contrast, noisy ultrasound data. The model leverages the nested skip connections of U-Net++ and the residual learning of ResNet-34 to achieve state-of-the-art segmentation accuracy. Evaluations of the developed model against the vast fetal ultrasound image collection yielded superior results by reaching 97.52% Dice coefficient as well as 95.15% Intersection over Union (IoU), and 3.91 mm Hausdorff distance. The pipeline integrated Grad-CAM++ allows explanations of the model decisions for clinical utility and trust enhancement. The explainability component enables medical professionals to study how the model functions, which creates clear and proven segmentation outputs for better overall reliability. The framework fills in the gap between AI automation and clinical interpretability by showing important areas which affect predictions. The research shows that deep learning combined with Explainable AI (XAI) operates to generate medical imaging solutions that achieve high accuracy. The proposed system demonstrates readiness for clinical workflows due to its ability to deliver a sophisticated prenatal diagnostic instrument that enhances healthcare results.

Craniofacial morphometrics in sharks provide crucial insights into evolutionary history, geographical variation, sexual dimorphism, and developmental patterns. However, the fragile cartilaginous nature of shark craniofacial skeleton poses significant challenges for traditional specimen preparation, often resulting in damaged cranial landmarks and compromised measurement accuracy. While computed tomography (CT) offers a non-invasive alternative for anatomical observation, the high electron density of dermal denticles in sharks creates a unique challenge, obstructing clear visualization of internal structures in three-dimensional volume-rendered images (3DVRI). This study presents an artificial intelligence (AI)-based solution using machine-learning algorithms for digitally removing dermal denticle layer from CT scans of shark craniofacial skeleton. We developed a geometric AI-driven software (SKINPEELER) that selectively removes high-intensity voxels corresponding to dermal denticle layer while preserving underlying anatomical structures. We evaluated this approach using CT scans from 20 sharks (16 Carcharhinus brachyurus, 2 Alopias vulpinus, 1 Sphyrna lewini, and 1 Prionace glauca), applying our AI-driven software to process the Digital Imaging and Communications in Medicine (DICOM) images. The processed scans were reconstructed using bone reconstruction algorithms to enable precise craniofacial measurements. We assessed the accuracy of our method by comparing measurements from the processed 3DVRIs with traditional manual measurements. The AI-assisted approach demonstrated high accuracy (86.16-98.52%) relative to manual measurements. Additionally, we evaluated reproducibility and repeatability using intraclass correlation coefficients (ICC), finding high reproducibility (ICC: 0.456-0.998) and repeatability (ICC: 0.985-1.000 for operator 1 and 0.882-0.999 for operator 2). Our results indicate that this AI-enhanced digital denticle removal technique, combined with 3D CT reconstruction, provides a reliable and non-destructive alternative to traditional specimen preparation methods for investigating shark craniofacial morphology. This novel approach enhances measurement precision while preserving specimen integrity, potentially advancing various aspects of shark research including evolutionary studies, conservation efforts, and anatomical investigations.