An efficient enhanced feature framework for grading of renal cell carcinoma using Histopathological Images

Maqsood, F ORCID: https://orcid.org/0009-0002-8723-9980, Wang, Z, Ali, MM, Qiu, B, Mahmood, T and Sarwar, R ORCID: https://orcid.org/0000-0002-0640-807X (2025) An efficient enhanced feature framework for grading of renal cell carcinoma using Histopathological Images. Applied Intelligence, 55 (2). 196. ISSN 0924-669X

Accepted Version File will be available on: 20 December 2025. Available under License In Copyright. Download (3MB)

Abstract

Renal cell carcinoma (RCC) represents the primary type of kidney cancer, responsible for approximately 85% of kidney cancer-related fatalities. Precise grading of this cancer is pivotal for tailoring effective treatments. Detecting RCC early, before metastasis, significantly improves survival rates. While Artificial intelligence-based classification methods have emerged for RCC, advancements in accuracy, processing efficiency, and memory utilization remain imperative. This study introduces the Efficient Enhanced Feature Framework (EFF-Net), a deep neural network architecture designed for RCC grading using histopathological image analysis. EFF-Net amalgamates potent feature extraction from convolutional layers with efficient Separable convolutional layers, aiming to accelerate model inference, reduce trainable parameters, mitigate overfitting, and elevate RCC grading precision. Evaluation across three distinct datasets showcases the EFF-Net's outstanding performance: achieving 91.90% accuracy, a precision of 91.4%, a recall of 91.8%, and a harmonic mean of precision and recall (F1 score) of 91.9% on the Kasturba Medical College (KMC) dataset. Additionally, on the Lung and Colon Dataset, EFF-Net achieved 99.8% accuracy, a precision of 99.7%, a recall of 99.9%, and a 98.7% F1 score. Similarly, the Acute Lymphoblastic Leukaemia dataset demonstrated remarkable performance: 99.8% accuracy, a precision of 99%, a recall of 99%, and a 99.7% F1 score. EFF-Net's superior accuracy surpasses existing state-of-the-art approaches while exhibiting reduced trainable parameters and computational requirements.