Use this link to cite:
https://hdl.handle.net/2183/49570 Representación y análisis de células mitóticas en histología H&E mediante autoencoders dispersos basados en LISTA
Loading...
Identifiers
Publication date
Authors
Pardo García, Enrique
Advisors
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen] La detección automática de figuras mitóticas en imágenes histopatológicas H&E constituye una tarea de gran relevancia clínica, pero limitada por el elevado coste de obtener anotaciones de calidad. Este trabajo explora si las representaciones dispersas aprendidas de forma no supervisada mediante la arquitectura LISTA son capaces de capturar información discriminativa entre figuras mitóticas y no mitóticas, sin emplear etiquetas durante el entrenamiento. Para ello se diseña e implementa un autoencoder disperso basado en un encoder LISTA y un decoder de pesos atados, evaluado sobre parches del dataset MIDOG en seis configuraciones experimentales que combinan tres funciones de pérdida de reconstrucción –MSE, MAE y SSIM– con dos dimensiones del diccionario, m = 128 y m = 1024. Los resultados muestran que, si bien las tres funciones de pérdida se comportan de forma equivalente en el régimen de diccionario reducido, su comportamiento diverge radicalmente al escalar el tamaño del código: MSE y MAE colapsan a una representación trivial en la práctica totalidad de su trayectoria de entrenamiento, mientras que SSIM mantiene la estabilidad e incluso mejora la calidad de reconstrucción. El análisis del potencial discriminativo revela una señal débil pero consistente, presente en átomos individuales del diccionario aunque no traducida en una separación visible del espacio de código. Estos hallazgos sitúan la elección de la función de pérdida y el tamaño del diccionario como los factores más determinantes de la viabilidad del enfoque propuesto.
[Abstract] Automatic detection of mitotic figures in H&E histopathology images is a task of high clinical relevance, but one constrained by the high cost of obtaining quality annotations. This work explores whether sparse representations learned in an unsupervised manner through the LISTA architecture are able to capture discriminative information between mitotic and non-mitotic figures, without using labels during training. To this end, a sparse autoencoder is designed and implemented, composed of a LISTA encoder and a tied-weights decoder, and evaluated on patches from the MIDOG dataset across six experimental configurations combining three reconstruction loss functions –MSE, MAE and SSIM– with two dictionary dimensions, m = 128 and m = 1024. The results show that, while the three loss functions behave equivalently under the smaller dictionary regime, their behaviour diverges sharply when the code size is scaled up: MSE and MAE collapse to a trivial representation over the vast majority of their training trajectory, whereas SSIM remains stable and even improves reconstruction quality. The analysis of discriminative potential reveals a weak but consistent signal, present in individual dictionary atoms yet not translated into a visible separation of the code space. These findings identify the choice of loss function and dictionary size as the most decisive factors for the viability of the proposed approach.
[Abstract] Automatic detection of mitotic figures in H&E histopathology images is a task of high clinical relevance, but one constrained by the high cost of obtaining quality annotations. This work explores whether sparse representations learned in an unsupervised manner through the LISTA architecture are able to capture discriminative information between mitotic and non-mitotic figures, without using labels during training. To this end, a sparse autoencoder is designed and implemented, composed of a LISTA encoder and a tied-weights decoder, and evaluated on patches from the MIDOG dataset across six experimental configurations combining three reconstruction loss functions –MSE, MAE and SSIM– with two dictionary dimensions, m = 128 and m = 1024. The results show that, while the three loss functions behave equivalently under the smaller dictionary regime, their behaviour diverges sharply when the code size is scaled up: MSE and MAE collapse to a trivial representation over the vast majority of their training trajectory, whereas SSIM remains stable and even improves reconstruction quality. The analysis of discriminative potential reveals a weak but consistent signal, present in individual dictionary atoms yet not translated into a visible separation of the code space. These findings identify the choice of loss function and dictionary size as the most decisive factors for the viability of the proposed approach.
Description
Keywords
Autoencoders dispersos Arquitectura LISTA Aprendizaje no supervisado Histología H&E Detección de figuras mitóticas Funciones de pérdida MSE, MAE y SSIM Codificación dispersa Sparse autoencoders LISTA architecture Unsupervised learning H&E histology Mitotic figure detection MSE, MAE and SSIM loss functions Sparse coding
Editor version
Rights
Attribution 4.0 International








