Ana Maria Marques da Silva

Precise annotations for large medical image datasets can be time-consuming. Additionally, when dealing with volumetric regions of interest, it is typical to apply segmentation techniques on 2D slices, compromising important information for accurately segmenting 3D structures. This study presents a deep learning pipeline that simultaneously tackles both challenges. Firstly, to streamline the annotation process, we employ a semi-automatic segmentation approach using bounding boxes as masks, which is less time-consuming than pixel-level delineation. Subsequently, recursive self-training is utilized to enhance annotation quality. Finally, a 2.5D segmentation technique is adopted, wherein a slice of a volumetric image is segmented using a pseudo-RGB image. The pipeline was applied to segment the carotid artery tree in T1-weighted brain magnetic resonance images. Utilizing 42 volumetric non-contrast T1-weighted brain scans from four datasets, we delineated bounding boxes around the carotid arteries in the axial slices. Pseudo-RGB images were generated from these slices, and recursive segmentation was conducted using a Res-Unet-based neural network architecture. The model’s performance was tested on a separate dataset, with ground truth annotations provided by a radiologist. After recursive training, we achieved an Intersection over Union (IoU) score of (0.68 ± 0.08) on the unseen dataset, demonstrating commendable qualitative results.

This study presents a semi-supervised learning strategy to predict Modified Radiographic Assessment of Lung Edema (mRALE) scores from chest radiographs (CXR) of COVID-19 patients. Using a combination of EfficientNet models and large volumes of unlabeled data, we developed a robust deep-learning model that exhibits outstanding performance, securing 4th place in the MIDRC mRALE Mastermind Challenge. Our strategy harnessed both the labeled data provided in the challenge and a significant amount of unlabeled data from the Medical Imaging and Data Resource Center (MIDRC) and CheXpert. We used labeled data to train an EfficientNet-B4 model, which we used to infer pseudo labels on our unlabeled data. We then used Chexpert to pre-train an EfficientNet-B7 model from scratch, allowing it to extract valuable information from a broader scope of CXR images. The final step involved fine-tuning the pre-trained EfficientNet-B7 model using additional CXR images from MIDRC. This approach effectively adapted the model to our specific domain of interest: portable CXR images from COVID-19 patients. The success of this semi-supervised learning approach highlights its potential to enhance diagnostic capabilities in medical imaging, particularly in situations where labeled data is limited. It demonstrates the feasibility of leveraging massive volumes of unlabeled data, even when the specific disease differs, provided that radiological characteristics are similar. This study paves the way for advancements in deep learning applications in healthcare, especially in emergency scenarios that demand swift and accurate diagnostics, such as the COVID-19 pandemic.

This study aims to evaluate non-invasive PET quantification methods for (R)-[¹¹C]PK11195 uptake measurement in multiple sclerosis (MS) patients and healthy controls (HC) in comparison with arterial input function (AIF) using dynamic (R)-[¹¹C]PK11195 PET and magnetic resonance images. The total volume of distribution (VT) and distribution volume ratio (DVR) were measured in the gray matter, white matter, caudate nucleus, putamen, pallidum, thalamus, cerebellum, and brainstem using AIF, the image-derived input function (IDIF) from the carotid arteries, and pseudo-reference regions from supervised clustering analysis (SVCA). Uptake differences between MS and HC groups were tested using statistical tests adjusted for age and sex, and correlations between the results from the different quantification methods were also analyzed. Significant DVR differences were observed in the gray matter, white matter, putamen, pallidum, thalamus, and brainstem of MS patients when compared to the HC group. Also, strong correlations were found in DVR values between non-invasive methods and AIF (0.928 for IDIF and 0.975 for SVCA, p < 0.0001). On the other hand, (R)-[¹¹C]PK11195 uptake could not be differentiated between MS patients and HC using VT values, and a weak correlation (0.356, p < 0.0001) was found between VT^AIF and VT^IDIF. Our study shows that the best alternative for AIF is using SVCA for reference region modeling, in addition to a cautious and appropriate methodology.

A tomografia por emissão de pósitrons (PET) e a ressonância magnética (MR) são modalidades relevantes no estudo do envelhecimento cerebral. A deposição de placa beta-amiloide (bA) pode ser avaliada por PET utilizando radiotraçadores específicos, como [¹¹C]PiB ou [¹⁸F]Florbetaben. A escala Centiloide (CL) foi desenvolvida para reduzir a variabilidade do protocolo de processamento e produzir uma quantificação harmonizada das imagens PET de bA. O objetivo deste estudo é apresentar o processo de implementação da escala CL para o radiotraçador [¹⁸F]Florbetaben em nossa instituição. Para tanto, é preciso validar o pipeline de processamento e análise das imagens com os dados disponibilizados pela Rede Interativa da Associação Global de Alzheimer (GAAIN) para o radiotraçador [¹⁸F]Florbetaben e [¹¹C]PiB. O GAAIN possui imagens de PET e MR de 10 controles (bA negativos) e 25 indivíduos bA positivos, incluindo idosos cognitivamente saudáveis, com comprometimento cognitivo leve, doença de Alzheimer e demência frontotemporal. Na escala CL são utilizados os valores de captação entre região cortical global (CTX), que considera as regiões córtex frontal, temporal, parietal, precuneus, estriato anterior e córtex insular, e o cerebelo inteiro (WC), como região de referência^. O processamento das imagens foi realizado utilizando o software PMOD, com a ferramenta PNEURO. Como resultado, obtivemos uma forte correlação entre os valores de [¹⁸F]Florbetaben e [¹¹C]PiB para os mesmos indivíduos. A equação final que transforma os valores SUVR do [¹⁸F]Florbetaben para escala CL em nossa instituição é: CL^FBB= 156,47 * SUVR*_PIBind - 157,47. A equação foi validada e atingiu os valores mínimos requeridos pelo Klunk et al (2015) de coeficiente de determinação R²=0,997, coeficiente angular igual a 1,00 e coeficiente linear igual a 1,34.Conclui-se que a implementação da escala CL em nossa instituição é viável,constituindo um passo importante para a utilização clínica dessa escala para investigação longitudinal de indivíduos em processo de envelhecimento usando radiotraçadores bA.

Interstitial Lung Disease (ILD) refers to pulmonary disorders that affect the lung parenchyma through inflammation and fibrosis. It is possible to diagnose ILD visually with computed tomography (CT), but it is highly demanding. Machine learning (ML) has yielded powerful models, such as convolutional neural networks (CNN), that achieve state-of-the-art performance in image classification. However, even with advances in CNN explainability, an expert is often required to justify its decisions adequately. Radiomic features are more reada ble for medical analysis because they can be related to image characteristics and are intuitively used by radiologists. There is potential in using image data via CNN and radiomic features to classify lung CT images. In this work, we develop two ML models: a CNN for classifying ILD using CT scans; and a Multi-Layer Perceptron (MLP) for classifying healthy and ILD using radiomic features. In the ensemble approach, output weights of each model are combined, providing a robust method capable of classifying ILD with the CT and the radiomic features. From a high-resolution CT dataset with 32 x 32 patches of pathological lung and healthy tissues, we extract 92 radiomic features, excluding those above 90% Pearson correlation in the training sets of both cross-validation and final models. Using 0.6 for the MLP and 0.4 for the CNN as weights, our approach achieves an accuracy of 0.874, while the MLP achieved 0.870 and, the CNN.

High-resolution structural magnetic resonance imaging (MRI) allows neurological investigation, especially when brain volumes must be carefully delineated to monitor neurodegeneration, such as in multiple sclerosis (MS). This study compares different segmentation techniques applied to brain MRI to measure the white matter (WM) and grey matter (GM) in healthy and MS brains. We propose to evaluate the reliability and how each segmentation method could potentially affect clinical trials in MS. Four segmentation software were evaluated: Statistical Parametric Mapping (SPM), Lesion Segmentation Tool (LST), Freesurfer, and Siena/X. We simulated healthy and MS brain MRI and compared the segmentation volumes with the ground truth. Our results showed that LST provides overall good segmentation with low variability. When SienaX spatially normalizes the images, the WM and GM volumes are overestimated. On the other hand, Freesurfer underestimates volumes. We conclude that the use of different segmentation software produces variability in GM and WM volumes, especially in challenging situations, such as small lesions and in the presence of noise. The best method was the automatic region growth algorithm implemented using the LST, which uses T1-weighted and T2-FLAIR MRI.

SARS-CoV-2 is the causative agent of COVID-19 and leaves characteristic impressions on chest Computed Tomography (CT) images in infected patients and this analysis is performed by radiologists through visual reading of lung images, and failures may occur. In this article, we propose a classification model, called Wavelet Convolutional Neural Network (WCNN) that aims to improve the differentiation of images of patients with COVID-19 from images of patients with other lung infections. The WCNN model was based on a Convolutional Neural Network (CNN) and wavelet transform. The model proposes a new input layer added to the neural network, which was called Wave layer. The hyperparameters values were defined by ablation tests. WCNN was applied to chest CT images to images from two internal and one external repositories. For all repositories, the average results of Accuracy (ACC), Sensitivity (Sen) and Specificity (Sp) were calculated. Subsequently, the average results of the repositories were consolidated, and the final values were ACC = 0.9819, Sen = 0.9783 and Sp = 0.98. The WCNN model uses a new Wave input layer, which standardizes the network input, without using data augmentation, resizing and segmentation techniques, maintaining the integrity of the tomographic image analysis. Thus, applications developed based on WCNN have the potential to assist radiologists with a second opinion in the analysis.1