Claudia Mazo

Nutritional deficiencies in coffee plants can reduce crop yield and quality, while their visual diagnosis requires specialised knowledge. This study evaluates a heterogeneous hierarchical transfer‐learning framework with hard routeing for classifying nine nutritional‐deficiency categories in coffee leaf images. The analysis used 1000 images from the CoLeaf dataset. Otsu thresholding was applied to isolate the leaf region, and four augmentation and class‐balancing configurations were evaluated using VGG19, ResNet50, MobileNetV2 and InceptionV3 with frozen ImageNet‐pretrained backbones. An initial direct classifier predicted the nine categories simultaneously. The hierarchical framework subsequently reformulated the task into four groups: group 1 comprised boron and calcium; group 2 comprised iron, magnesium and manganese; group 3 comprised nitrogen, phosphorus and potassium; and group 4 represented multiple deficiencies. Images assigned to groups 1–3 were routed to the corresponding specialist classifier, whereas images assigned to group 4 were directly classified as multiple deficiencies. For the nine‐class direct baseline, MobileNetV2‐E3 obtained an accuracy of 63.39%, a macro‐F1 of 58.21% and a weighted‐F1 of 62.98%. The selected round 1 group‐level classifier, MobileNetV2‐E2, obtained 77.63%, 65.76% and 75.63%, respectively. Conditional evaluation within the corresponding group‐specific subsets produced weighted‐F1 values of 97.44% for the specialist VGG19‐E2 of group 1, 75.78% for the specialist MobileNetV2‐E3 of group 2 and 84.10% for the specialist MobileNetV2‐E4 of group 3. Nitrogen and multiple deficiencies remained challenging, obtaining class‐specific F1‐scores of 68.75% and 13.33%, respectively. The findings demonstrate the feasibility of selecting different architectures and data configurations for individual hierarchical nodes. Independent external validation and evaluation with propagated routeing errors remain necessary.

Vision-language models and their adaptations to image segmentation tasks present enormous potential for producing highly accurate and interpretable results. However, implementations based on CLIP and BiomedCLIP are still lagging behind more sophisticated architectures such as CRIS. In this work, instead of focusing on text prompt engineering as is the norm, we attempt to narrow this gap by showing how to ensemble vision-language segmentation models (VLSMs) with a low-complexity CNN. By doing so, we achieve a significant Dice score improvement of 6.3% on the BKAI polyp dataset using the ensembled BiomedCLIPSeg, while other datasets exhibit gains ranging from 1% to 6%. Furthermore, we provide initial results on additional four radiology and non-radiology datasets. We conclude that ensembling works differently across these datasets (from outperforming to underperforming the CRIS model), indicating a topic for future investigation by the community. The code is available at https://github.com/juliadietlmeier/VLSM-Ensemble.

Background: Cerebral small vessel disease (CSVD) is a common neurological condition that contributes to strokes, dementia, disability, and mortality worldwide. We conducted a systematic review and meta-analysis to investigate the use of neuroimaging CSVD markers in machine learning (ML) based diagnosis and prognosis of cognitive impairment and dementia, and identify both methodological changes over time and barriers to clinical translation. Methods: Following the PRISMA guidelines, we systematically searched for original studies that used both neuroimaging CSVD markers and ML methods for diagnosing and prognosing neurodegenerative diseases (preregistration in PROSPERO: CRD42022366767). Each paper was independently reviewed by a pair of reviewers at all stages, with a third consulted to resolve conflicts. We meta-analysed the effectiveness of ML models to distinguish healthy controls from Alzheimer’s dementia and cognitive impairment, using area under the curve (AUC) as the performance metric. Results: We identified 75 studies: 43 on diagnosis, 27 on prognosis, and 5 on both. Nearly 60% of studies were published in the past two years, reflecting a growing interest in using CSVD markers in ML-based diagnosis and prognosis of neurodegenerative diseases, especially Alzheimer’s dementia. This rising interest may be linked to the strong performance of such models: according to our meta-analysis, ML approaches using CSVD markers perform well in differentiating healthy controls from Alzheimer’s dementia (AUC 0.88 [95%-CI 0.85–0.92]) and cognitive impairment (AUC 0.84 [95%-CI 0.74–0.95]). However, the growing interest has not been matched by methodological rigour: only 16 studies met the criteria for inclusion in the meta-analysis due to inconsistent reporting, only five assessed the generalisability of their models on external datasets, and six lacked clear diagnostic criteria. Conclusions: Interest in incorporating CSVD markers into ML models for neurodegenerative disease classification is on the rise, and their performance suggests that this is worth further exploration. Serious methodological issues, including inconsistent reporting, limited generalisability testing, and other potential biases, are unfortunately common and hinder further adoption. Our targeted recommendations provide a roadmap to accelerate the integration of ML into clinical practice.

Brain Stroke is a life-threatening medical condition that requires prompt diagnosis and treatment to minimize its devastating effects like partial paralysis, speech impairment, and memory loss. The classification of brain strokes is crucial for accurate and timely medical treatment. Our research employs transfer learning, such as Convolutional Neural Networks (CNN), ResNet, RadiNet, and MobileNet algorithms, to classify brain strokes from Computed Tomography (CT) scans, focusing on improving both the speed and accuracy of stroke diagnosis using transfer learning techniques. To deal with the unbalanced nature of the dataset, we implemented various features of data augmentation techniques and achieved class balance in the dataset. Our proposed model MobileNet on augmented down-sampling dataset demonstrates, exceptional overall F1-scores achieving up to 99%, precision of 99% and recall of 99 % for no brain stroke cases and F1-score of 99%, precision of 99%, and recall of 100% for stroke cases, with an overall accuracy of 99 %. We conclude that our research adds a valuable and efficient method in classifying a CT scan into a stroke or no brain stroke case, which is still an open problem despite the usage of deep learning and machine learning techniques.

Epilepsy is a neurological disorder characterized by spontaneous recurrent seizures. While 20% to 30% of epilepsy cases are untreatable with Anti-Epileptic Drugs, some of these cases can be addressed through surgical intervention. The success of such interventions greatly depends on accurately locating the epileptogenic tissue, a task achieved using diagnostic techniques like Stereotactic Electroencephalography (SEEG). SEEG utilizes multi-modal fusion to aid in electrode localization, using pre-surgical resonance and post-surgical computer tomography images as inputs. To ensure the absence of artifacts or misregistrations in the resultant images, a fusion method that accounts for electrode presence is required. We proposed an image fusion method in SEEG that incorporates electrode segmentation from computed tomography as a sampling mask during registration to address the fusion problem in SEEG. The method was validated using eight image pairs from the Retrospective Image Registration Evaluation Project (RIRE). After establishing a reference registration for the MRI and identifying eight points, we assessed the method’s efficacy by comparing the Euclidean distances between these reference points and those derived using registration with a sampling mask. The results showed that the proposed method yielded a similar average error to the registration without a sampling mask, but reduced the dispersion of the error, with a standard deviation of 0.86 when a mask was used and 5.25 when no mask was used.

Although it is widely assumed that Artificial Intelligence (AI) will revolutionise healthcare in the near future, considerable progress must yet be made in order to gain the trust of healthcare professionals and patients. Improving AI transparency is a promising avenue for addressing such trust issues. However, transparency still lacks maturation and definitions. We seek to answer what challenges do experts and professionals in computing and healthcare identify concerning transparency of AI in healthcare? Here, we examine AI transparency in healthcare from five angles: interpretability, privacy, security, equity, and intellectual property. We respond to this question based on recent literature discussing the transparency of AI in healthcare and on an international online survey we sent to professionals working in computing and healthcare and potentially within AI. We collected responses from 40 professionals around the world. Overall, the survey results and current state of the art suggest key problems are a generalised lack of information available to the general public, a lack of understanding of transparency aspects covered in this work, and a lack of involvement of all stakeholders in the development of AI systems. We propose a set of recommendations, the implementation of which can enhance the transparency of AI in healthcare.