Table 1.

A brief comparison between previous studies that proposed techniques for prostate histopathology images.

Reference	Study Aim	Year	Strength	Weakness	Number of Patients
[2]	Automated classification using AdaBoost-based Ensemble Learning	2016	They integrated various feature descriptors, different color channels, and classifiers.	The algorithm able to discover only the critical regions on the digital slides	50
[14]	A novel technique of labeling individual glands as malignant or benign was proposed.	2013	The technique can detect individual malignant gland units without relying on the neighboring histology and/or the spatial extent of the cancer.	It applied on a small number of radical prostatectomy patients	8
[15]	Methodology for automated gland and nuclei segmentation	2008	They incorporate low-, high-level knowledge, and structural constraints imposed via domain knowledge.	They focused on a smaller cohort of cancer images and the dataset is private	44
[16]	A new automated method for gland segmentation	2017	This method texture- and gland structure-based methods	The method failed in the images with the cribriform pattern. They validated data using 2-fold cross validation	10
[17]	Multistage Segmentation Using Sample Entropy Texture Analysis	2020	An added advantage of performing multistage segmentation using sample entropy values is that one could easily separate epithelial nuclei from the stroma nuclei in standard H&E stained images without using any additional immunohistochemical (IHC) markers.	It requires identifying sample entropy features	25
[18]	A new approach to identify prostate cancer areas in complex	2014	It utilizes the differential information embedded in the intensity characteristics of H&E images to quickly classify areas of the prostate tissue	Classification performance is tested using only KNN algorithm	20
[19]	Ensemble based system for feature selection and classification	2011	They addressed the possibility of missing tumor regions through the use of tile-based probabilities and heat maps.	They focused only on texture feature selection and not used a voting schema for the ensemble classifier to enhance the probability scores	14
[20]	A novel fully automated CAD system	2006	The proposed system represents the first attempt to automatically analyse histopathology across multiple scales	Their system trained using only 3 images	6
[21]	A new multiclass approach	2018	It obtained improved grading results	It was evaluated based on its impact on the performance of the ensemble framework only	213
[22]	A bag-of-words approach to classify images using SpeededUp Robust Features (SURF)	2016	The drawbacks of scale-invariant feature transform descriptor is overcome by the SURF descriptors causing an enhanced output accuracy	More features needed to be integrated with their feature extraction process to enhance accuracy of the classification	75
[23]	An automatic method for segmentation and classification (Integration of Salp Swarm Optimization Algorithm and Rider Optimization Algorithm)	2019	Less time complexity	The maximal accuracy, sensitivity, and specificity does not exceed 90%	20
[24]	A new region-based convolutional neural network framework for multi-task prediction	2018	The model achieved a detection accuracy 99.07% with an average area under the curve of 0.998	They didn’t have patient-level information with which to perform a more rigorous patient-level stratification.	40
[25]	An approach to nuclei segmentation using a conditional generative adversarial network	2019	It enforces higher-order consistency and captures better results when compared to conventional CNN models.	The model trained on small annotated patches	34
[26]	Deep neural network algorithm for segmentation of individual nuclei	2019	A simple, fast, and parameter-free postprocessing procedure is done to get the final segmented nuclei as one 1000 × 1000 image can be segmented in less than 5 s.	The model is trained on a small number of images and has been tested on the images that may have different appearances	30
[27]	Two novel approaches (combination of 4 types of feature descriptors, advanced machine-learning classifiers) to automatically identify prostate cancer	2019	They apply for the first time on prostate segmented glands, deep-learning algorithms modifying the popular VGG19 neural network.	The hand-driven learning approach employs SVM, where selecting the suitable kernel function could be tricky	35
[28]	Automated Gleason grading via deep learning	2018	The study showed promising results especially for cases with heterogeneous Gleason patterns	The model trained on small mini patches at each iteration	886
[29]	A deep learning system using the U-Net	2019	The system outperformed 10 out of 15 pathologists	The system was built upon three pretrained preprocessing modules, each of which still required pixel-wise annotations.	1243
[30]	Predicting Gleason Score Using OverFeat Trained Deep CNN as feature extractor	2016	It is quite effective, even without from-scratch training on WSI tiles. Processing time is low	Small size of patches	213
[31]	CNN to idiomatically identify the features	2016	The system is not constrained to H&E stained images and could easily be applied to immunohistochemistry	Some detection errors happen at the boundaries of the tissue	254
[32]	DL model to detect cancer based on NASNetLarge architecture and high-quality annotated training dataset	2020	The model demonstrated its strong ability in prediction as accuracy attained 98%	The availability of fully digitalized cohorts represents a bottleneck	400
[33]	A novel benchmark was designed for measuring and comparing the performances of different CNN models with the proposed PROMETEO	2021	Average processing time is less compared to other architectures	The network validated on 3-fold cross-validation method	470
[34]	Novel features that include spatial inter-nuclei statistics and intra-nuclei properties for discriminating high-grade prostate cancer patterns	2018	The system tackled the inter-observer variability in prostate grading and can lead to a consensus-based training that improves both classification	lack examples of the highest grades of disease	56