Enhancing head and neck tumor management with artificial intelligence: Integration and perspectives

Seminars in Cancer Biology, Journal Year: 2023, Volume and Issue: 95, P. 52 - 74

Published: July 18, 2023

Language: Английский

---

Enhancing the reliability of deep learning-based head and neck tumour segmentation using uncertainty estimation with multi-modal images

Physics in Medicine and Biology, Journal Year: 2024, Volume and Issue: 69(16), P. 165018 - 165018

Published: July 26, 2024

Abstract Objective. Deep learning shows promise in autosegmentation of head and neck cancer (HNC) primary tumours (GTV-T) nodal metastases (GTV-N). However, errors such as including non-tumour regions or missing still occur. Conventional methods often make overconfident predictions, compromising reliability. Incorporating uncertainty estimation, which provides calibrated confidence intervals can address this issue. Our aim was to investigate the efficacy various estimation improving segmentation We evaluated their levels voxel predictions ability reveal potential errors. Approach. retrospectively collected data from 567 HNC patients with diverse sites multi-modality images (CT, PET, T1-, T2-weighted MRI) along clinical GTV-T/N delineations. Using nnUNet 3D pipeline, we compared seven methods, evaluating them based on accuracy (Dice similarity coefficient, DSC), calibration (Expected Calibration Error, ECE), (Uncertainty-Error overlap using DSC, UE-DSC). Main results. Evaluated hold-out test dataset ( n = 97), median DSC scores for GTV-T GTV-N across all had a narrow range, 0.73 0.76 0.78 0.80, respectively. In contrast, ECE exhibited wider 0.30 0.12 0.25 0.09 GTV-N. Similarly, UE-DSC also ranged broadly, 0.21 0.38 0.22 0.36 A probabilistic network—PhiSeg method consistently demonstrated best performance terms UE-DSC. Significance. study highlights importance enhancing reliability deep GTV. The results show that while be similar reliability, measured by error uncertainty-error overlap, varies significantly. Used visualisation maps, these may effectively pinpoint uncertainties at level.

Language: Английский

---

Addressing fairness issues in deep learning-based medical image analysis: a systematic review

npj Digital Medicine, Journal Year: 2024, Volume and Issue: 7(1)

Published: Oct. 17, 2024

Deep learning algorithms have demonstrated remarkable efficacy in various medical image analysis (MedIA) applications. However, recent research highlights a performance disparity these when applied to specific subgroups, such as exhibiting poorer predictive elderly females. Addressing this fairness issue has become collaborative effort involving AI scientists and clinicians seeking understand its origins develop solutions for mitigation within MedIA. In survey, we thoroughly examine the current advancements addressing issues MedIA, focusing on methodological approaches. We introduce basics of group subsequently categorize studies fair MedIA into evaluation unfairness mitigation. Detailed methods employed are presented too. Our survey concludes with discussion existing challenges opportunities establishing healthcare system. By offering comprehensive review, aim foster shared understanding among researchers clinicians, enhance development methods, contribute creation an equitable society.

Language: Английский

---

Bayesian Uncertainty Estimation by Hamiltonian Monte Carlo: Applications to Cardiac MRI Segmentation

The Journal of Machine Learning for Biomedical Imaging, Journal Year: 2024, Volume and Issue: 2(June 2024), P. 856 - 887

Published: June 23, 2024

Deep learning (DL)-based methods have achieved state-of-the-art performance for many medical image segmentation tasks. Nevertheless, recent studies show that deep neural net- works (DNNs) can be miscalibrated and overconfident, leading to ”silent failures” are risky clinical applications. Bayesian DL provides an intuitive approach failure de- tection, based on posterior probability estimation. However, the is intractable large DNNs. To tackle this challenge, we propose a framework using Hamiltonian Monte Carlo (HMC), tempered by cold (CP) accommodate data augmentation, named HMC-CP. For HMC compu- tation, further cyclical annealing strategy, capturing both local global geometries of distribution, enabling highly efficient DNN training with same computational budget as single DNN. The resulting outputs ensemble along uncertainty. We evaluate proposed HMC-CP extensively cardiac magnetic resonance (MRI) segmentation, in-domain steady-state free precession (SSFP) cine images well out-of-domain datasets quantitative T1 T2 mapping. Our results method improves accuracy uncertainty estimation in- data, compared well-established baseline such Dropout Ensembles. Additionally, establish conceptual link between commonly known stochastic gradient descent (SGD) provide general insight into DL. This implicitly encoded in dynamics but often overlooked. With reliable estimation, our promising direction toward trustworthy release code <a href='https://gitlab.tudelft.nl/yidongzhao/hmc_uncertainty'>https://gitlab.tudelft.nl/yidongzhao/hmc_uncertainty</a>

Language: Английский

---

Ensemble Deep Learning Models for Automated Segmentation of Tumor and Lymph Node Volumes in Head and Neck Cancer Using Pre- and Mid-Treatment MRI: Application of Auto3DSeg and SegResNet

Lecture notes in computer science, Journal Year: 2025, Volume and Issue: unknown, P. 259 - 273

Published: Jan. 1, 2025

Language: Английский

---

Uncertainty-Aware Deep Learning for Segmentation of Primary Tumor and Pathologic Lymph Nodes in Oropharyngeal Cancer: Insights from a Multi-Center Cohort

Computerized Medical Imaging and Graphics, Journal Year: 2025, Volume and Issue: unknown, P. 102535 - 102535

Published: March 1, 2025

Language: Английский

---

Application of artificial intelligence in head and neck tumor segmentation: A comparative systematic review and meta-analysis between PET and PET/CT modalities

Research Square (Research Square), Journal Year: 2025, Volume and Issue: unknown

Published: May 21, 2025

Background For the effective treatment planning of head and neck cancers, precise tumor segmentation is vital. The combination artificial intelligence (AI) technology with imaging systems like positron emission tomography (PET) PET/ computed (PET/CT) has made attempts to automate these processes. Despite attempts, usefulness AI PET compared PET/CT still lacks clarity. Methods A comprehensive search was performed on Scopus, Embase, PubMed, Cochrane, Web Science, Google Scholar for studies published before Dec 2024, an update in March 2025. Included utilized algorithms segment tumors via or provided quantitative performance measures. Pooled estimates Dice Similarity Coefficient (DSC) sensitivity, precision, Hausdorff Distance (HD95) were calculated using a random-effects model. Also, sensitivity analyses find potential source heterogeneity. Additionally, subgroup conducted overall primary segmentation. Publication bias assessed weighted Egger’s test, followed by presentation funnel plots different metrics. Risk (RoB) evaluated QUADAS-C tool. CLAIM used assess methodological quality robustness included studies. Results Eleven included. All rated as having low risk bias. scores showed high There significant difference between PET-only modalities. effectiveness metrics improvement their respective DSC 0.05 (95% CI: 0.033–0.071), precision ~ 0.04 each, HD95 decreased approximately 3 mm. heterogeneity most except HD95, which (I2 = 75%. In it shown that one study caused heterogeneity, which, after its exclusion, 67.5%. analyses, two groups, including segmentation, did not show differences metric. Conclusions AI-assisted greater than tumors. These results justify clinical use AI-based beyond contouring due automation highlight importance unified datasets alongside distributed learning improve applicability consistency workflows. The protocol registered at PROSPERO [CRD42024614436].

Language: Английский

---

Integration of Radiomics and Tumor Biomarkers in Interpretable Machine Learning Models

Cancers, Journal Year: 2023, Volume and Issue: 15(9), P. 2459 - 2459

Published: April 25, 2023

Despite the unprecedented performance of deep neural networks (DNNs) in computer vision, their practical application diagnosis and prognosis cancer using medical imaging has been limited. One critical challenges for integrating diagnostic DNNs into radiological oncological applications is lack interpretability, preventing clinicians from understanding model predictions. Therefore, we study propose integration expert-derived radiomics DNN-predicted biomarkers interpretable classifiers which call ConRad, computerized tomography (CT) scans lung cancer. Importantly, tumor are predicted a concept bottleneck (CBM) such that once trained, our ConRad models do not require labor-intensive time-consuming biomarkers. In evaluation application, only input to segmented CT scan. The proposed compared convolutional (CNNs) act as black box classifier. We further investigated evaluated all combinations radiomics, CNN features five different classifiers. found non-linear SVM logistic regression with Lasso outperform others five-fold cross-validation, although highlight interpretability its primary advantage. used feature selection, substantially reduces number non-zero weights while increasing accuracy. Overall, combines CBM-derived an ML perform excellently nodule malignancy classification.

Language: Английский

---

Development of Clinical Radiomics-Based Models to Predict Survival Outcome in Pancreatic Ductal Adenocarcinoma: A Multicenter Retrospective Study

Diagnostics, Journal Year: 2024, Volume and Issue: 14(7), P. 712 - 712

Published: March 28, 2024

Purpose. This multicenter retrospective study aims to identify reliable clinical and radiomic features build machine learning models that predict progression-free survival (PFS) overall (OS) in pancreatic ductal adenocarcinoma (PDAC) patients. Methods. Between 2010 2020 pre-treatment contrast-enhanced CT scans of 287 pathology-confirmed PDAC patients from two sites the Hopital Universitaire de Bruxelles (HUB) 47 hospitals within HUB network were retrospectively analysed. Demographic, clinical, data also collected. Gross tumour volume (GTV) non-tumoral pancreas (RPV) semi-manually segmented radiomics extracted. Patients comprised training dataset, while those remaining constituted testing dataset. A three-step method was used for feature selection. Based on GradientBoostingSurvivalAnalysis classifier, different trained tested OS PFS. Model performances assessed using C-index Kaplan–Meier curves. SHAP analysis applied allow post hoc interpretability. Results. total 107 extracted each GTV RPV. Fourteen subgroups selected: GTV, RPV, & GTV-volume RPV-volume both Subsequently, 14 Gradient Boosting Survival Analysis tested. In model demonstrated highest performance (C-index: 0.72) among all other models, PFS, exhibited a superior 0.70). Conclusions. An integrated approach, combining features, excels predicting OS, whereas demonstrate strong PFS prediction.

Language: Английский

---

Ground Truth from Multiple Manually Marked Images to Evaluate Blood Vessel Segmentation

Lecture notes in electrical engineering, Journal Year: 2024, Volume and Issue: unknown, P. 531 - 536

Published: Jan. 1, 2024

Language: Английский

---