Motivation and objectives – Early cancer detection is essential for improving patient outcomes, and 18F-FDG PET/CT imaging plays a central role by combining metabolic and anatomical information. However, accurate lesion detection remains challenging due to the presence of multiple lesions with varying sizes and locations. This study investigates whether incorporating anatomical prior information can improve deep learning-based lesion detection performance. Methods – Anatomical priors were incorporated by adding organ segmentation masks generated with TotalSegmentator as auxiliary input channels to two lesion detection frameworks: the CNN-based nnDetection and a transformer-based Swin UNETR implemented in MONAI. The Swin Transformer was trained using a two-stage strategy, with self-supervised pretraining performed on the autoPET dataset and supervised fine-tuning of the detector model conducted on the independent Karolinska lymphoma dataset. Model evaluation followed a single hold-out split, and performance was assessed using FROC and average precision metrics. Results – Experiments were conducted on two independent PET/CT datasets covering different tracers and cancer subtypes. The autoPET dataset includes 18F-FDG PET/CT scans of lymphoma, melanoma, and lung cancer, while the Karolinska dataset focuses on lymphoma imaging. Incorporating anatomical priors consistently improved lesion detection performance within the nnDetection framework across both datasets. Specifically, nnDetection augmented with anatomical masks improved in mAP@0.1–0.5 from 0.288 to 0.335. In contrast, anatomical priors had minimal impact on the Swin Transformer, which did not demonstrate clear advantages over CNN-based encoders. Conclusions – Anatomy-aware priors substantially enhance lesion detection performance in CNN-based models, highlighting the importance of explicit anatomical context for multi-lesion PET/CT analysis. However, these benefits do not readily transfer to transformer-based architectures, indicating the need for improved strategies to integrate anatomical information into vision transformers for medical image analysis.

Accurate and robust segmentation of heterogeneous brain tumors is critical for individualized treatment planning, yet the integration of diverse multicenter datasets remains challenging due to patient privacy constraints. The BraTS Challenge Generalizability Task (GoAT) has highlighted the importance of developing models that generalize across multiple tumor subtypes, including adult glioma, meningioma, and brain metastasis, along with pediatric and sub-Saharan cohorts. In this work, we present BraTS-FL, a federated learning (FL) approach integrated with the nnU-Net framework to collaboratively train segmentation models across distinct tumor subtypes without sharing raw data between institutions. We design a multi-client FL setup, with each client specializing in a specific tumor subtype and employing harmonized preprocessing and training via the MONet bundle. Comparative evaluation on the BraTS 2025 generalizability validation set demonstrates that BraTS-FL achieves competitive performance compared to centralized nnU-Net training in terms of Dice and surface-based metrics across all tumor subregions. These findings underscore FL’s viability for privacy-preserving, scalable, and generalizable brain tumor segmentation in real-world heterogeneous clinical settings.

Artificial intelligence (AI) is driving major changes across numerous fields, with healthcare emerging as one of the areas with the greatest potential for impact. In medical imaging, AI has the potential to enhance patient care through personalized treatment planning and early disease detection, while simultaneously supporting clinicians by optimizing their workload, and automating complex tasks such as radiological image analysis. Over the past decade, substantial progress has been made in medical AI research, leading to highly accurate and robust models in controlled experimental settings. However, bringing these AI tools into everyday clinical use has proven challenging. Despite many scientific breakthroughs, only few AI systems are currently being adopted in clinical settings. 

This PhD thesis focuses on understanding why that gap exists, and how to bridge it. The work explores the technical, organizational, and ethical barriers that slow down AI adoption in healthcare, and proposes new ways to make AI more practical, transparent, and trustworthy in clinical environments.

A key result of this research is MAIA, a collaborative platform designed to bring together doctors, radiologists, and AI researchers. MAIA provides a shared space where experts can jointly develop and test AI tools under realistic clinical conditions. By combining research methods with everyday medical workflows, MAIA helps accelerate the transition from experimental AI models to clinical tools. The platform has been successfully deployed in both research and hospital environments, demonstrating its effectiveness in accelerating the integration of AI into medical practice.

Building on this foundation, the thesis also introduces MONet, a framework that makes it easier to adapt and reuse state-of-the-art medical image segmentation models for different healthcare applications. It enables smooth integration of AI into various clinical settings, from federated learning across different institutions to human-in-the-loop smart annotation tools, ensuring that research innovations can be efficiently transferred into real-world practice.

Finally, as a methodological contribution to the field, the thesis investigates the incorporation of anatomical and contextual prior knowledge into existing deep learning frameworks, with the goal of improving model interpretability and anatomical awareness. These methods were evaluated on different clinical tasks, such as lung lobe segmentation on chest CT, breast cancer treatment response prediction, and lymphoma segmentation on whole-body PET/CT, with the findings suggesting that the relevance of anatomical priors is task-dependent and can vary significantly across contexts.

In summary, the thesis work aims to contribute to bridging the gap between AI research and clinical implementation by developing collaborative infrastructures, adaptable frameworks, and methodological insights that support the trustworthy, transparent, and effective integration of AI technologies in medical imaging practice.

Artificiell intelligens (AI) medför stora förändringar inom många områden, där hälso- och sjukvården framstår som ett av dem med störst potential. Inom medicinsk bildbehandling kan AI förbättra patientvården genom tidig detektion av sjukdom och individualiserad planering av behandling, samtidigt som den utgör ett stöd för kliniker genom att minska arbetsbelastningen och automatisera komplexa uppgifter, såsom analys av radiologiska bilder. Under det senaste decenniet har betydande framsteg gjorts inom medicinsk AI-forskning, vilket har lett till mycket noggranna och robusta modeller i kontrollerade experimentella miljöer. 

Att införa dessa AI-verktyg i den dagliga kliniska praktiken har dock visat sig vara en utmaning. Trots många vetenskapliga genombrott används fortfarande endast ett fåtal AI-system inom klinisk rutin. Denna doktorsavhandling fokuserar på att förstå varför denna klyfta existerar och hur den kan överbryggas. Arbetet undersöker tekniska, organisatoriska och etiska hinder som fördröjer införandet av AI i vården, och föreslår nya strategier för att göra AI mer praktisk, transparent och pålitlig i kliniska miljöer. 

Ett centralt resultat av denna forskning är MAIA, en plattform utformad för samarbete mellan kliniska läkare, radiologer och AI-forskare. MAIA erbjuder en gemensam miljö där experter kan utveckla och testa AI-verktyg under realistiska kliniska förhållanden. Genom att kombinera forskningsmetoder med vardagliga kliniska arbetsflöden bidrar MAIA till att påskynda överföringen från experimentella AI-modeller till kliniska verktyg. Plattformen har framgångsrikt implementerats både i forsknings- och sjukhusmiljöer, vilket visar att den effektivt kan integrera AI i medicinsk praxis. 

Med denna plattform som grund presenterar avhandlingen också MONet, ett ramverk som underlättar anpassning och återanvändning av avancerade modeller för medicinsk bildsegmentering i olika vårdsammanhang. Ramverket möjliggör smidig integration av AI i olika kliniska miljöer, från federerat lärande mellan olika institutioner till smarta annoteringsverktyg baserade på en människa i loopen, vilket säkerställer att forskningsinnovationer effektivt kan överföras till verklig klinisk praktik. 

Som ett metodologiskt bidrag undersöker avhandlingen dessutom hur anatomisk och kontextuell förkunskap kan införlivas i befintliga modeller för djupinlärning, med målet att förbättra modellernas tolkbarhet och anatomiska medvetenhet. Dessa metoder har utvärderats på olika kliniska uppgifter, såsom lunglobssegmentering på bröstkorgs-CT, prediktion av behandlingssvar vid bröstcancer och lymfomsegmentering på helkropps-PET/CT. Resultaten visar att betydelsen av anatomiska förkunskaper är uppgiftsberoende och kan variera avsevärt mellan olika kliniska kontexter. 

Sammanfattningsvis syftar denna avhandling till att bidra till att minska klyftan mellan AI-forskning och klinisk tillämpning genom utveckling av infrastruktur för samarbete, anpassningsbara ramverk och metodologiska insikter som stödjer en pålitlig, transparent och effektiv integration av AI-teknologier inom medicinsk bildbehandling.

The widespread success of nnU-Net as a state-of-the-art tool for medical image segmentation has driven its adoption as a baseline, but its limited portability and lack of clinical integration have limited broader deployment in real-world healthcare workflows. To address these challenges, we present the MONet Bundle, extending nnU-Net within the MONAI ecosystem, providing a modular benchmarking tool for Federated Learning (FL) that is directly compatible with downstream clinical operations such as model deployment, active learning, and DICOM-based PACS integration. MONet enables federated training across distributed clinical datasets while maintaining standardized preprocessing and harmonized workflows. Its flexibility is validated on two representative segmentation tasks: lymphoma lesion segmentation in PET-CT and brain tumor segmentation from the BraTS challenge. In both settings, MONet’s federated models consistently outperformed cross-site baselines and approached, or in some cases outperformed, the performance of centralized task-fusion models with minimal user intervention. The code is available at https://github.com/SimoneBendazzoli93/MONet-Bundle.

Radiation therapy is a primary and effective treatment strategy for NasoPharyngeal Carcinoma (NPC). The precise delineation of Gross Tumor Volumes (GTVs) and Organs-At-Risk (OARs) is crucial in radiation treatment, directly impacting patient prognosis. Despite that deep learning has achieved remarkable performance on various medical image segmentation tasks, its performance on OARs and GTVs of NPC is still limited, and high-quality benchmark datasets on this task are highly desirable for model development and evaluation. To alleviate this problem, the SegRap2023 challenge was organized in conjunction with MICCAI2023 and presented a large-scale benchmark for OAR and GTV segmentation with 400 Computed Tomography (CT) scans from 200 NPC patients, each with a pair of pre-aligned non-contrast and contrast-enhanced CT scans. The challenge aimed to segment 45 OARs and 2 GTVs from the paired CT scans per patient, and received 10 and 11 complete submissions for the two tasks, respectively. In this paper, we detail the challenge and analyze the solutions of all participants. The average Dice similarity coefficient scores for all submissions ranged from 76.68% to 86.70%, and 70.42% to 73.44% for OARs and GTVs, respectively. We conclude that the segmentation of relatively large OARs is well-addressed, and more efforts are needed for GTVs and small or thin OARs. The benchmark remains available at: https://segrap2023.grand-challenge.org.

Purpose: Glaucoma leads to pathological loss of axons in the retinal nerve fibre layer at the optic nerve head (ONH). This study aimed to develop a strategy for the estimation of the cross‐sectional area of the axons in the ONH. Furthermore, improving the estimation of the thickness of the nerve fibre layer, as compared to a method previously published by us.

Methods: In the 3D‐OCT image of the ONH, the central limit of the pigment epithelium and the inner limit of the retina, respectively, were identified with deep learning algorithms. The minimal distance was estimated at equidistant angles around the circumference of the ONH. The cross‐sectional area was estimated by the computational algorithm. The computational algorithm was applied on 16 non‐glaucomatous subjects.

Results: The mean cross‐sectional area of the waist of the nerve fibre layer in the ONH was 1.97 ± 0.19 mm². The mean difference in minimal thickness of the waist of the nerve fibre layer between our previous and the current strategies was estimated as CIμ (0.95) 0 ± 1 μm (d.f. = 15).

Conclusions: The developed algorithm demonstrated an undulating cross‐sectional area of the nerve fibre layer at the ONH. Compared to studies using radial scans, our algorithm resulted in slightly higher values for cross‐sectional area, taking the undulations of the nerve fibre layer at the ONH into account. The new algorithm for estimation of the thickness of the waist of the nerve fibre layer in the ONH yielded estimates of the same order as our previous algorithm.

Purpose Our study investigates the potential benefits of incorporating prior anatomical knowledge into a deep learning (DL) method designed for the automated segmentation of lung lobes in chest CT scans. Approach We introduce an automated DL-based approach that leverages anatomical information from the lung's vascular system to guide and enhance the segmentation process. This involves utilizing a lung vessel connectivity (LVC) map, which encodes relevant lung vessel anatomical data. Our study explores the performance of three different neural network architectures within the nnU-Net framework: a standalone U-Net, a multitasking U-Net, and a cascade U-Net. Results Experimental findings suggest that the inclusion of LVC information in the DL model can lead to improved segmentation accuracy, particularly, in the challenging boundary regions of expiration chest CT volumes. Furthermore, our study demonstrates the potential for LVC to enhance the model's generalization capabilities. Finally, the method's robustness is evaluated through the segmentation of lung lobes in 10 cases of COVID-19, demonstrating its applicability in the presence of pulmonary diseases. Conclusions Incorporating prior anatomical information, such as LVC, into the DL model shows promise for enhancing segmentation performance, particularly in the boundary regions. However, the extent of this improvement has limitations, prompting further exploration of its practical applicability.

Objectives Quantitative CT imaging is an important emphysema biomarker, especially in smoking cohorts, but does not always correlate to radiologists' visual CT assessments. The objectives were to develop and validate a neural network-based slice-wise whole-lung emphysema score (SWES) for chest CT, to validate SWES on unseen CT data, and to compare SWES with a conventional quantitative CT method. Materials and methods Separate cohorts were used for algorithm development and validation. For validation, thin-slice CT stacks from 474 participants in the prospective cross-sectional Swedish CArdioPulmonary bioImage Study (SCAPIS) were included, 395 randomly selected and 79 from an emphysema cohort. Spirometry (FEV1/FVC) and radiologists' visual emphysema scores (sum-visual) obtained at inclusion in SCAPIS were used as reference tests. SWES was compared with a commercially available quantitative emphysema scoring method (LAV950) using Pearson's correlation coefficients and receiver operating characteristics (ROC) analysis. Results SWES correlated more strongly with the visual scores than LAV950 (r=0.78 vs. r=0.41, p<0.001). The area under the ROC curve for the prediction of airway obstruction was larger for SWES than for LAV950 (0.76 vs. 0.61, p=0.007). SWES correlated more strongly with FEV1/FVC than either LAV950 or sum-visual in the full cohort (r=-0.69 vs. r=-0.49/r=-0.64, p<0.001/p=0.007), in the emphysema cohort (r=-0.77 vs. r=-0.69/r=-0.65, p=0.03/p=0.002), and in the random sample (r=-0.39 vs. r=-0.26/r=-0.25, p=0.001/p=0.007). ConclusionT he slice-wise whole-lung emphysema score (SWES) correlates better than LAV950 with radiologists' visual emphysema scores and correlates better with airway obstruction than do LAV950 and radiologists' visual scores. Clinical relevance statementThe slice-wise whole-lung emphysema score provides quantitative emphysema information for CT imaging that avoids the disadvantages of threshold-based scores and is correlated more strongly with reference tests than LAV950 and reader visual scores.

Automated detection of tumour lesions on positron emission tomography–computed tomography (PET/CT) image data is a clinically relevant but highly challenging task. Progress in this field has been hampered in the past owing to the lack of publicly available annotated data and limited availability of platforms for inter-institutional collaboration. Here we describe the results of the autoPET challenge, a biomedical image analysis challenge aimed to motivate research in the field of automated PET/CT image analysis. The challenge task was the automated segmentation of metabolically active tumour lesions on whole-body ¹⁸F-fluorodeoxyglucose PET/CT. Challenge participants had access to a large publicly available annotated PET/CT dataset for algorithm training. All algorithms submitted to the final challenge phase were based on deep learning methods, mostly using three-dimensional U-Net architectures. Submitted algorithms were evaluated on a private test set composed of 150 PET/CT studies from two institutions. An ensemble model of the highest-ranking algorithms achieved favourable performance compared with individual algorithms. Algorithm performance was dependent on the quality and quantity of data and on algorithm design choices, such as tailored post-processing of predicted segmentations. Future iterations of this challenge will focus on generalization and clinical translation.

Significant advances have been made toward building accurate automatic segmentation models for adult gliomas. However, the performance of these models often degrades when applied to pediatric glioma due to their imaging and clinical differences (domain shift). Obtaining sufficient annotated data for pediatric glioma is typically difficult because of its rare nature. Also, manual annotations are scarce and expensive. In this work, we propose Domain-Adapted nnU-Net (DA-nnUNet) to perform unsupervised domain adaptation from adult glioma (source domain) to pediatric glioma (target domain). Specifically, we add a domain classifier connected with a gradient reversal layer (GRL) to a backbone nnU-Net. Once the classifier reaches a very high accuracy, the GRL is activated with the goal of transferring domain-invariant features from the classifier to the segmentation model while preserving segmentation accuracy on the source domain. The accuracy of the classifier slowly degrades to chance levels. No annotations are used in the target domain. The method is compared to 8 different supervised models using BraTS-Adult glioma (N=1251) and BraTS-PED glioma data (N=99). The proposed method shows notable performance enhancements in the tumor core (TC) region compared to the model that only uses adult data: ~32% better Dice scores and ~20 better 95th percentile Hausdorff distances. Moreover, our unsupervised approach shows no statistically significant difference compared to the practical upper bound model using manual annotations from both datasets in TC region. The code is shared at https://github.com/Fjr9516/DA_nnUNet.