Cataracts, known for lens clouding and being a common cause of visual impairment, persist as a primary contributor to vision loss and blindness, presenting notable diagnostic and prognostic challenges. This work presents a novel framework called the Cataract States Detection Network (CSDNet), which utilizes deep learning methods to improve the detection of cataract states. The aim is to create a framework that is more lightweight and adaptable for use in environments or devices with limited memory or storage capacity. This involves reducing the number of trainable parameters while still allowing for effective learning of representations from data. Additionally, the framework is designed to be suitable for real-time or near-real-time applications where rapid inference is essential. This study utilizes cataract and normal images from the Ocular Disease Intelligent Recognition (ODIR) database. The suggested model employs smaller kernels, fewer training parameters, and layers to efficiently decrease the number of trainable parameters, thereby lowering computational costs and average running time compared to other pre-trained models such as VGG19, ResNet50, DenseNet201, MIRNet, Inception V3, Xception, and Efficient net B0. The experimental results illustrate that the proposed approach achieves a binary classification accuracy of 97.24% (normal or cataract) and an average cataract state detection accuracy of 98.17% (normal, grade 1-minimal cloudiness, grade 2-immature cataract, grade 3-mature cataract, and grade 4-hyper mature cataract), competing with state-of-the-art cataract detection methods. The resulting model is lightweight at 17 MB and has fewer trainable parameters (175, 617), making it suitable for deployment in environments or devices with constrained memory or storage capacity. With a runtime of 212 ms, it is well-suited for real-time or near-real-time applications requiring rapid inference.

Breast cancer is the second most common cancer among women worldwide, and the diagnosis by pathologists is a time-consuming procedure and subjective. Computer-aided diagnosis frameworks are utilized to relieve pathologist workload by classifying the data automatically, in which deep convolutional neural networks (CNNs) are effective solutions. The features extracted from the activation layer of pre-trained CNNs are called deep convolutional activation features (DeCAF). In this paper, we have analyzed that all DeCAF features are not necessarily led to higher accuracy in the classification task and dimension reduction plays an important role. We have proposed reduced DeCAF (R-DeCAF) for this purpose, and different dimension reduction methods are applied to achieve an effective combination of features by capturing the essence of DeCAF features. This framework uses pre-trained CNNs such as AlexNet, VGG-16, and VGG-19 as feature extractors in transfer learning mode. The DeCAF features are extracted from the first fully connected layer of the mentioned CNNs, and a support vector machine is used for classification. Among linear and nonlinear dimensionality reduction algorithms, linear approaches such as principal component analysis (PCA) represent a better combination among deep features and lead to higher accuracy in the classification task using a small number of features considering a specific amount of cumulative explained variance (CEV) of features. The proposed method is validated using experimental BreakHis and ICIAR datasets. Comprehensive results show improvement in the classification accuracy up to 4.3% with a feature vector size (FVS) of 23 and CEV equal to 0.15.