A Dual-branch Framework Based on Implicit Continuous Representation for Tumor Image Segmentation DOI Creative Commons
Jing Wang, Yuanjie Zheng, Junxia Wang

et al.

Research Square (Research Square), Journal Year: 2023, Volume and Issue: unknown

Published: Nov. 7, 2023

Abstract Breast tumor segmentation has important significance for early detection and determination of treatment plans. However, segmenting early-stage small tumors in breast images is challenging due to low-resolution regions, variation shapes, blurred boundaries. More importantly, scans are usually noisy include metal artifacts. Most the existing methods have difficulty extracting lesion discriminative information, leading problem that ignored or predictions contain a lot noise. In addition, common reconstruction algorithms based on discrete ignore continuity feature space. Therefore, this paper, we investigate novel flexible dual-branch framework, named High-Resolution Information Bottleneck-based Segmentation Network (HR-IBS), segmentation. For first time, method introduces high-resolution region (HR-TR) branch via implicit neural representations learning functions map input signal continuous density. The enables from regions another branch. Furthermore, design an bottleneck-based (IBS) branch, which adopts information bottleneck U-Net retain features most relevant while removing discovering more informative regions. branches interact with each other facilitate performance. Comprehensive experiments conducted benchmarks two modalities images. results show proposed outperforms models contributes optimizing hand-crafted ground-truths.

Language: Английский

PET/CT based transformer model for multi-outcome prediction in oropharyngeal cancer DOI
Baoqiang Ma, Jiapan Guo, Alessia de Biase

et al.

Radiotherapy and Oncology, Journal Year: 2024, Volume and Issue: 197, P. 110368 - 110368

Published: June 2, 2024

Language: Английский

Citations

8
Deep learning-based outcome prediction using PET/CT and automatically predicted probability maps of primary tumor in patients with oropharyngeal cancer DOI Creative Commons
Alessia de Biase, Baoqiang Ma, Jiapan Guo

et al.

Computer Methods and Programs in Biomedicine, Journal Year: 2023, Volume and Issue: 244, P. 107939 - 107939

Published: Nov. 22, 2023

Recently, deep learning (DL) algorithms showed to be promising in predicting outcomes such as distant metastasis-free survival (DMFS) and overall (OS) using pre-treatment imaging head neck cancer. Gross Tumor Volume of the primary tumor (GTVp) segmentation is used an additional channel input DL improve model performance. However, binary mask GTVp directs focus network defined region only uniformly. models trained for have also been generate predicted probability maps (TPM) where each pixel value corresponds degree certainty that classified tumor. The aim this study was explore effect TPM extra CT- PET-based prediction oropharyngeal cancer (OPC) patients terms local control (LC), regional (RC), DMFS OS.

Language: Английский

Citations

16
Artificial intelligence uncertainty quantification in radiotherapy applications − A scoping review DOI Creative Commons
Kareem A. Wahid, Zaphanlene Kaffey, David P. Farris

et al.

Radiotherapy and Oncology, Journal Year: 2024, Volume and Issue: 201, P. 110542 - 110542

Published: Sept. 17, 2024

Language: Английский

Citations

6
Application of simultaneous uncertainty quantification and segmentation for oropharyngeal cancer use-case with Bayesian deep learning DOI Creative Commons
Jaakko Sahlsten, Joel Jaskari, Kareem A. Wahid

et al.

Communications Medicine, Journal Year: 2024, Volume and Issue: 4(1)

Published: June 8, 2024

Abstract Background Radiotherapy is a core treatment modality for oropharyngeal cancer (OPC), where the primary gross tumor volume (GTVp) manually segmented with high interobserver variability. This calls reliable and trustworthy automated tools in clinician workflow. Therefore, accurate uncertainty quantification its downstream utilization critical. Methods Here we propose uncertainty-aware deep learning OPC GTVp segmentation, illustrate utility of multiple applications. We examine two Bayesian (BDL) models eight measures, utilize large multi-institute dataset 292 PET/CT scans to systematically analyze our approach. Results show that uncertainty-based approach accurately predicts quality segmentation 86.6% cases, identifies low performance cases semi-automated correction, visualizes regions segmentations likely fail. Conclusions Our BDL-based analysis provides first-step towards more widespread implementation segmentation.

Language: Английский

Citations

5
Enhancing the reliability of deep learning-based head and neck tumour segmentation using uncertainty estimation with multi-modal images DOI Creative Commons
Jintao Ren, Jonas Teuwen, Jasper Nijkamp

et al.

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

Citations

4
Recommendations for the creation of benchmark datasets for reproducible artificial intelligence in radiology DOI Creative Commons
Nikos Sourlos, Rozemarijn Vliegenthart, João Santinha

et al.

Insights into Imaging, Journal Year: 2024, Volume and Issue: 15(1)

Published: Oct. 14, 2024

Various healthcare domains have witnessed successful preliminary implementation of artificial intelligence (AI) solutions, including radiology, though limited generalizability hinders their widespread adoption. Currently, most research groups and industry access to the data needed for external validation studies. The creation accessibility benchmark datasets validate such solutions represents a critical step towards generalizability, which an array aspects ranging from preprocessing regulatory issues biostatistical principles come into play. In this article, authors provide recommendations in explain current limitations realm, explore potential new approaches. CLINICAL RELEVANCE STATEMENT: Benchmark datasets, facilitating AI software performance can contribute adoption clinical practice. KEY POINTS: are essential performance. Factors like image quality representativeness cases should be considered. help by increasing trustworthiness robustness AI.

Language: Английский

Citations

4
Gross tumor volume confidence maps prediction for soft tissue sarcomas from multi-modality medical images using a diffusion model DOI Creative Commons
Yafei Dong, Thibault Marin, Yue Zhuo

et al.

Physics and Imaging in Radiation Oncology, Journal Year: 2025, Volume and Issue: 33, P. 100734 - 100734

Published: Jan. 1, 2025

Language: Английский

Citations

0
Uncertainty-Aware Deep Learning for Segmentation of Primary Tumor and Pathologic Lymph Nodes in Oropharyngeal Cancer: Insights from a Multi-Center Cohort DOI Creative Commons
Alessia de Biase, Nanna M. Sijtsema, Lisanne V. van Dijk

et al.

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

Published: March 1, 2025

Language: Английский

Citations

0
Harnessing uncertainty in radiotherapy auto-segmentation quality assurance DOI Creative Commons
Kareem A. Wahid, Jaakko Sahlsten, Joel Jaskari

et al.

Physics and Imaging in Radiation Oncology, Journal Year: 2023, Volume and Issue: 29, P. 100526 - 100526

Published: Dec. 19, 2023

We have read with great interest the article by Outeiral et al. [1Rodríguez R. Ferreira Silvério N. González P.J. Schaake E.E. Janssen T. van der Heide U.A. al.A network score-based metric to optimize quality assurance of automatic radiotherapy target segmentations.Phys Imaging Radiat Oncol. 2023; 28100500PubMed Google Scholar], in which authors propose a simple for optimizing deep learning (DL) auto-segmentation workflows. commend their insightful analysis using two meticulously curated MRI datasets. Through this study, probed into relatively underexplored yet clinically relevant domain uncertainty estimation auto-segmentation. One key contributions study is reappropriation standard DL outputs as indicator identify cases that clinicians should review further. The achieve applying an empirically derived threshold softmax output network, computing mean thresholded score map (termed HiS metric), and correlating it geometric indices. When juxtaposed entropy — commonly used measure model consistently demonstrated stronger correlation indices, suggesting its superior ability stratify needing additional review. applaud authors' efforts novel would like note some potential caveats could pave way future research directions. Conventional large networks often yield overconfident predictions can result poor calibration [2Guo C, Pleiss G, Sun Y, Weinberger KQ. On modern neural networks. In: Precup D, Teh YW, editors. Proceedings 34th international conference on machine learning, vol. 70, PMLR; 06--11 Aug 2017, p. 1321–30.Google meaning predicted probabilities do not align true underlying data. This discrepancy undermine reliability these detecting out-of-distribution data, critical aspect systems. Notably, direct use measures point contention within community [3Holm A.N. Wright D. Augenstein I. Revisiting approximation text classification.Information. 14: 420Crossref Scopus (2) Scholar, 4Pearce T, Brintrup A, Zhu J. Understanding confidence uncertainty. arXiv [csLG] 2021.Google Scholar]. However, moderately sized exhibit well-calibrated performance [5Jaskari Sahlsten Damoulas Knoblauch Särkkä S. Kärkkäinen L. al.Uncertainty-aware methods robust diabetic retinopathy classification.IEEE Access. 2022; 10: 76669-76681Crossref (10) Therefore, unclear whether had major impact al.'s ​​nnU-Net architecture. In contrast, Bayesian approaches been observed be may circumvent issues [6Izmailov P, Vikram S, Hoffman MD, Wilson AGG. What are posteriors really like? Meila M, Zhang 38th 139, 18--24 Jul 2021, 4629–40.Google Specifically, application approximate techniques, such Monte Carlo dropout [7Gal Ghahramani Z. Dropout approximation: representing learning. Balcan MF, KQ, 33rd 48, New York, USA: 20--22 Jun 2016, 1050–9.Google Scholar] or ensembles [8Lakshminarayanan B. Pritzel A. Blundell C. Simple scalable predictive ensembles.Adv Neural Inf Process Syst. 2017; 30Google (Fig. 1), compared conventional solutions. While demand slightly higher computational cost, they considered investigating studies. Importantly, ensembling (e.g., through cross-validation schemes) becoming increasingly common many solutions [9Eisenmann Reinke Weru V, Tizabi Isensee F, Adler TJ, Why winner best? 2023 IEEE/CVF Conference computer vision pattern recognition (CVPR), IEEE; 2023, 19955–66.Google previously benchmarked under U-net framework oropharyngeal cancer shown efficacy [10Sahlsten Jaskari Wahid K.A. Ahmed Glerean E. He al.Application simultaneous quantification image segmentation probabilistic learning: Performance benchmarking delineation use-case.medRxiv. https://doi.org/10.1101/2023.02.20.23286188Crossref (0) Interestingly, robustness analysis; merging ensemble improved when employing metric. Of note, alternative allow calibrated estimates, conformal prediction [11Angelopoulos AN, Bates A gentle introduction distribution-free quantification. also show promise further investigated. Finally, we proposed metric, if uncertainty, unable disentangle epistemic (i.e., intrinsic uncertainty) aleatoric extrinsic statistical [12Hüllermeier Waegeman W. Aleatoric concepts methods.Mach Learn. 2021; 110: 457-506Crossref (562) same said general entropy, there exist entropy-related metrics, expected mutual information, distinguish source combined approach 13Band N, Rudner TGJ, Feng Q, Filos Nado Z, Dusenberry MW, Benchmarking detection tasks. [statML] 2022.Google Moreover, distribution parameters assumed delta distribution, e.g., implicitly non-existent. depending specific case, combinations more suitable. An increasing number studies begun apply radiotherapy-related 14van den Berg C.A.T. Meliadò E.F. Uncertainty assessment applications.Semin 32: 304-318Crossref PubMed (7) 15Tang P. Yang Nie Wu X. Zhou Wang Y. Unified medical from end-to-end manner.Knowl-Based 241108215Crossref (60) 16De Biase Sijtsema N.M. Dijk Langendijk J.A. Ooijen P.M.A. Deep aided adaptive thresholding tumor probability FDG PET CT images.Phys Med Biol. https://doi.org/10.1088/1361-6560/acb9cfCrossref (6) 17Balagopal Nguyen Morgan H. Weng Dohopolski M. Lin M.-H. learning-based segmenting invisible clinical volumes estimated uncertainties post-operative prostate radiotherapy.Med Image Anal. 72102101Abstract Full Text PDF (31) 18Li Bagher-Ebadian Gardner Kim Elshaikh Movsas al.An uncertainty-aware architecture outlier mitigation gland treatment planning.Med Phys. 50: 311-322Crossref 19Bragman FJS, Tanno R, Eaton-Rosen Li W, Hawkes DJ, Ourselin multitask joint representations MR-only planning. Medical assisted intervention – MICCAI 2018, Springer International Publishing; 3–11.Google 20Chen Men K. Chen Tang al.CNN-based breast radiotherapy.Front 2020; 524Crossref (42) 21Cubero Serrano Castelli De Crevoisier Acosta O. Exploring al.IEEE 20th symposium biomedical imaging (ISBI).IEEE. 2023: 1-4Google 22Min Dowling Jameson M.G. Cloak Faustino Sidhom al.Clinical volume MRI-guided estimation.Radiother 186109794Abstract 23Lei Mei Ye Gu al.Automatic organs-at-risk head-and-neck separable convolutional hard-region-weighted loss.Neurocomputing. 442: 184-199Crossref (17) serves cornerstone contribution crucial literature. eagerly anticipate advances significant field work. During preparation work, ChatGPT (GPT-4 architecture; October 17, Version) improve grammatical accuracy semantic structure portions text. After tool, reviewed edited content needed take full responsibility publication. Kareem supported NCI NRSA Guided Cancer Therapy Training Program (T32CA261856). work Joel Jaskari, Jaakko Sahlsten, Kimmo Kaski was part Academy Finland Project 345449. Antti Mäkitie grant Finnish Society Sciences Letters. Benjamin Kann NIH/National Institute Dental Craniofacial Research (NIDCR) K08 Grant (K08DE030216). Clifton Fuller receives related support (T32CA261856), well unrelated salary/effort NIH institutes. infrastructure MD Anderson Center via: Charles Daneen Stiefel Head Neck Oropharyngeal Program; Image-guided Therapy; NIH/NCI Support (CCSG) Radiation Oncology (P30CA016672). David Fuentes R01CA195524.

Language: Английский

Citations

9
Probability maps for deep learning-based head and neck tumor segmentation: Graphical User Interface design and test DOI Creative Commons
Alessia de Biase, Liv Ziegfeld, Nanna M. Sijtsema

et al.

Computers in Biology and Medicine, Journal Year: 2024, Volume and Issue: 177, P. 108675 - 108675

Published: May 28, 2024

The different tumor appearance of head and neck cancer across imaging modalities, scanners, acquisition parameters accounts for the highly subjective nature manual segmentation task. variability contours is one causes lack generalizability suboptimal performance deep learning (DL) based auto-segmentation models. Therefore, a DL-based method was developed that outputs predicted probabilities each PET-CT voxel in form probability map instead fixed contour. aim this study to show DL-generated maps are clinically relevant, intuitive, more suitable solution assist radiation oncologists gross volume on images patients.

Language: Английский

Citations

3