Introduction

We have developed a lightweight and fast deep learning model called Light-Bladder-Net (LBN), which enhances its generalization ability through data augmentation with image transformation and the addition of uniform noise.

We also compared the performance using Weighted Voting and feature selection methods, employing model weighting and voting among multiple models to predict bladder cancer classification and enhance recognition capability.

The results showed that Weighted Voting performed the best, achieving an accuracy of 0.83, with a sensitivity of 0.85, specificity of 0.80, and precision of 0.81.

Workflow

Initially, missing values are filtered out, followed by data type conversion and image transformation.The analysis involves training models using both DL and ML approaches.

Subsequently, features are extracted from the DL models, and a second round of model training is conducted after feature screening.

Weighted voting is utilized, assigning different weights to multiple models, and the results are subsequently compared.

Light-Bladder-Net (LBN)

The architecture of the LBN model developed in this study, which consists of two 3×3 convolution layers with ReLU activation functions, two 2×2 maximum pooling layers, a normalization layer, a flatten layer, and a FC layer for the final classification result.

The LBN model offers the advantage of high-speed computing and low memory requirements, enabling more effective extraction of features from images