Juan Manuel Estrada Soubran

Mr. Juan Manuel Estrada Soubran

orcid

Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
5 September 2025 in Applied Sciences
Verified
Journal Article
Open access
PDF file
JAJulio Fernando Hernandez Angeles
Julio Fernando Hernandez Angeles
Alberto RosalesAlberto Rosales
Alberto Rosales
JKJean Marie Vianney Kinani
Jean Marie Vianney Kinani
JPJuan Pablo Francisco Posadas-Durán
Juan Pablo Francisco Posadas-Durán
Francisco Gallegos-FunesFrancisco Gallegos-Funes
Francisco Gallegos-Funes
JSJuan Manuel Estrada Soubran
Juan Manuel Estrada Soubran
Environmental pollution poses an increasing threat to public health, particularly in urban areas with high levels of pollutant exposure. To address this challenge, this study proposes a model based on fine-tuning the LLaMA 3 large language model for the classification of pollution-related diseases using user-reported symptoms. A balanced dataset was employed, with examples evenly distributed across 10 common diseases, and several preprocessing techniques were applied, including tokenization, normalization, noise removal, and data augmentation. The model was fine-tuned using the QLoRA technique, which integrates quantization with low-rank adaptation, enabling both training and inference on resource-constrained hardware. During training, a consistent reduction in loss and a progressive improvement in validation accuracy were observed. Moreover, the confusion matrix demonstrated a high classification success rate with minimal misclassification across classes. The findings suggest that optimized large language models can be effectively applied in settings with limited computational infrastructure, supporting the early diagnosis of diseases associated with environmental factors.

6 August 2026
Verified
Book Chapter
Open access
PDF file
JSJuan Manuel Estrada Soubran
Juan Manuel Estrada Soubran
Francisco Gallegos-FunesFrancisco Gallegos-Funes
Francisco Gallegos-Funes
EVErick Velazquez-Lozada
Erick Velazquez-Lozada
Alberto RosalesAlberto Rosales
Alberto Rosales
GUGuillermo Urriolagoitia-Sosa
Guillermo Urriolagoitia-Sosa
BRBeatriz Romero-Angeles
Beatriz Romero-Angeles
A binary semantic segmentation of nuclei cells in hematoxylin and eosin (H&E)-stained histopathological images is a fundamental yet challenging task in computational pathology, primarily due to significant color variability. This study presents a comprehensive evaluation of the impact of various color spaces on the performance of a U-Net model for this binary segmentation task. It is systematically trained and assessed the model using images represented in red–green–blue (RGB); grayscale; hue–saturation–value (HSV); hue–lightness–saturation (HLS); lightness, green–red opponent, blue–yellow opponent (CIE lαβ); lightness, u*, v* (LUV); and luminance, red-difference chrominance, blue-difference chrominance (YCrCb) color spaces; coupled with different optimizers (stochastic gradient descent [SGD], Adam, AdamW, NAdam). A robust multicriteria decision-making (MCDM) framework, employing the D-CRITIC method, was utilized for model evaluation. Our results demonstrate that nonconventional color spaces, particularly HSV with the AdamW optimizer, yield superior segmentation accuracy, achieving a DICE coefficient of 0.844 and an Intersection over Union (IoU) of 0.731. Furthermore, driven by these findings, it is proposed that a preprocessing pipeline (Denoising Vanilla Autoencoder with Contrast Enhancement [DVACE] + SwinIR‑based Contrast Limited Adaptive Histogram Equalization [SCLA]) leverages color spaces such as HSV, HLS, LUV, CIE lαβ, or YCrCb to construct enhanced RGB images. For example, when evaluated on a U-Net model trained from scratch with the AdamW optimizer using the YCrCb color space as a base, this pipeline achieved a DICE coefficient of 0.821 and an IoU of 0.697, substantially outperforming the baseline model trained on original RGB images without preprocessing (DICE: 0.802, IoU: 0.671). The study confirms that deliberate color space selection and targeted preprocessing are critical for developing robust and accurate deep learning (DL) solutions in digital pathology (DG).