Alexandre Henrique Neto

Abstract
Early detection of gastrointestinal lesions such as intestinal metaplasia (IM), dysplasia, and polyps remains challenging due to their subtle appearance and the scarcity of well-annotated medical image datasets. To address this limitation, we introduce Cut Instance Mixing (CIM), a domain-specific data augmentation method designed to generate anatomically plausible lesion-containing images through the identification of biologically relevant regions of interest and seamless lesion blending using Poisson image editing and gradient-based mixing. CIM was evaluated across three distinct endoscopic datasets (IM, dysplasia, and polyps) using a ResNet50 classifier and five-fold cross-validation. The proposed method consistently outperformed state-of-the-art augmentation techniques. In IM classification, CIM with alpha = 0.8 achieved the highest performance (AUC: 0.879, Accuracy: 0.823), surpassing MixUp, CutMix and random copy-paste. In dysplasia detection, CIM reached near-perfect results (AUC: 0.997, Accuracy: 0.966), and demonstrated strong generalization on an external polyp dataset (AUC: 0.830, Accuracy: 0.769). Grad-CAM analyses further confirmed that CIM preserves clinically relevant features, improving model attention on lesion regions. These findings demonstrate that CIM enables the generation of realistic and biologically coherent synthetic samples, effectively mitigating data imbalance and enhancing classification robustness. The method is architecture-agnostic and broadly applicable to tasks requiring anatomically consistent augmentation, providing a promising direction for improving deep learning systems in gastrointestinal imaging.

Abstract
Metaplasia detection in upper gastrointestinal endoscopy is crucial to identify patients at higher risk of gastric cancer. Deep learning algorithms can be useful for detecting and localising these lesions during an endoscopy exam. However, to train these types of models, a lot of annotated data is needed, which can be a problem in the medical field. To overcome this, data augmentation techniques are commonly applied to increase the dataset's variability but need to be adapted to the specificities of the application scenario. In this study, we discuss the potential benefits and identify four key research challenges of a promising data augmentation approach, namely image combination methodologies, such as CutMix, for metaplasia detection and localisation in gastric endoscopy imaging modalities.

Abstract
Gastric cancer (GC) is still a significant public health issue, among the most common and deadly cancers globally. The identification and characterization of precancerous lesions of the stomach using endoscopy are crucial for determining the risk of cancer and guiding appropriate surveillance. In this scenario, deep learning (DL)-based computer vision methods have the potential to help us classify and identify particular patterns in endoscopic images, leading to a more accurate classification of these types of lesions. The quantity and quality of the data used heavily influence the classification performance of DL networks. However, one of the major setbacks for developing high-performance DL classification models is the typical need for more available data in the medical field. This review explores the use of Generative Adversarial Networks (GANs) and classical data augmentation techniques for improving the classification of precancerous stomach lesions. GANs are DL models that have shown promising results in generating synthetic data, which can be used to augment limited medical datasets. This review discusses recent studies that have implemented GANs and classical data augmentation methods to improve the accuracy of cancerous lesion classification. The results indicate that GANs can effectively increase the dataset's size, enhance the classification models' performance. In specific applications, such as the augmentation of endoscopic images depicting gastrointestinal polyps and Barrett's esophagus Adenocarcinoma, our review reveals instances where GANs, including models like Deep Convolutional GANs and conditional GANs, outperform classical data augmentation methods. Furthermore, this review highlights the challenges and limitations of the recent works using GANs and classical data augmentation techniques in medical imaging analysis and proposes directions for future research.

Abstract
Gastric Cancer (GC) and Colorectal Cancer (CRC) are some of the most common cancers in the world. The most common diagnostic methods are upper endoscopy and biopsy. Possible expert distractions can lead to late diagnosis. GC is a less studied malignancy than CRC, leading to scarce public data that difficult the use of AI detection methods, unlike CRC where public data are available. Considering that CRC endoscopic images present some similarities with GC, a CRC Transfer Learning approach could be used to improve AI GC detectors. This paper evaluates a novel Transfer Learning approach for real-time GC detection, using a YOLOv4 model pre-trained on CRC detection. The results achieved are promising since GC detection improved relatively to the traditional Transfer Learning strategy.

Abstract
Colorectal cancer is the third most common cancer and the second cause of cancer-related deaths in the world. Colonoscopic surveillance is extremely important to find cancer precursors such as adenomas or serrated polyps. Identifying small or flat polyps can be challenging during colonoscopy and highly dependent on the colonoscopist's skills. Deep learning algorithms can enable improvement of polyp detection rate and consequently assist to reduce physician subjectiveness and operation errors. This study aims to compare YOLO object detection architecture with self-attention models. In this study, the Kvasir-SEG polyp dataset, composed of 1000 colonoscopy annotated still images, were used to train (700 images) and validate (300images) the performance of polyp detection algorithms. Well-defined architectures such as YOLOv4 and different YOLOv5 models were compared with more recent algorithms that rely on self-attention mechanisms, namely the DETR model, to understand which technique can be more helpful and reliable in clinical practice. In the end, the YOLOv5 proved to be the model achieving better results for polyp detection with 0.81 mAP, however, the DETR had 0.80 mAP proving to have the potential of reaching similar performances when compared to more well-established architectures.