Grivkov et al. [19] implemented InceptionNetV3 [20] for diagnosis of TB in Shenzhen (SH) and Montgomery (MG) datasets [21] and achieved 86.8% accuracy. Das et al. [22] exploited transfer learning to improve InceptionV3 accuracy to 91.7% on the same datasets. Priya et al. [23] implemented transfer learning on VGG19 [24], ResNet50 [25], DenseNet121 [26] and InceptionV3 models. In their analysis, pre-trained VGG19 has achieved 89% and 95% best accuracies on MG and SH datasets, respectively. Cao et al. [27] implemented DenseNet121, VGG and ResNet152 [25] models and achieved best accuracy of 90.38% classification accuracy with DenseNet121. Raman et al. [28] adopted somewhat different approach than the aforementioned methods. They have used three pre-trained models (ResNet101 [25], VGG19, and DenseNet201 [26]) to extract features from CXR images and use eXtreme Gradient Boosting (XG-Boost) (1.6.1, Tianqi Chen and Carlos Ernesto Guestrin, Seattle, WA, USA) [29] model to classify TB and non-TB in them. In their experiments, DenseNet201 with XG-Boost architecture achieved the highest accuracy of 99.92% as compared to its counterparts. Munadi et al. [30] proposed to enhance CXR quality before feeding them to pre-trained ResNet and EfficientNet [31] models. They have used three different image-enhancing techniques (unsharped masking, high-frequency emphasis filtering, and contrast limited adaptive histogram equalization). In their analysis, EfficientNet with unsharped masking image enhancement achieved 89.92% accuracy on SH dataset. Msnoda et al. [32] implemented ResNet, GoogLeNet [33], and AlexNet [34] with an extra Spatial Pyramid Pooling (SPP) [35] layer. Among the implemented architectures, GoogLeNet achieved the highest classification accuracy of 97%, which was then improved to 98% by using the SPP layer.

The methods discussed so far rely on the architecture of an individual model for classification efficiency. There are methods that combine two or even more models to form an ensemble network to achieve better performance. For example, Rajaraman et al. [36] implemented VGG16, InceptionResNetV2 [37], InceptionV3, XceptionNet [38] and DenseNet121, and then ranked them based on their accuracy. In their experiments, the top-3 models were InceptionV3 (accuracy = 94%), DenseNet121 (accuracy = 92.8%) and InceptionResNetV2 (accuracy = 92.5%). They have evaluated multiple ensemble methods to combine the top-3 models such as majority voting, simple averaging, weighted averaging stacking and blending to make an ensemble network. Their analysis showed that stacking ensemble demonstrated better performance and achieved 94.1% accuracy. Dasanayaka et al. [39] have implemented an ensemble of only two models (VGG16 and InceptionV3), and achieved 97.10% accuracy, which is higher than the ensemble of the three models proposed in [36]. Oloko-Oba et al. [40] have implemented an ensemble of VGG16, ResNet50 and InceptionV3 and achieved best accuracy of 96.14%. In their other study [41], they have explored ensemble of EfficientNets [31] for the diagnosis of TB. In their analysis of individual models, EfficientNet-B4 achieved best accuracy of 94.35% on SH dataset, which was then improved to 97.44% through ensemble learning. The ensemble was built by averaging the performance of three best individual EfficientNets (B2, B3, and B4). Saif et al. [42] proposed to combine the traditional hand-engineered feature with an ensemble of DenseNet169, ResNet50 and InceptionV3 models. Their ensemble model has achieved best accuracy of 99.7% on SH dataset. Overall, ensemble methods have shown superior performance for TB screening in CXR images than the individual models.

The PE algorithm is block-based and performs four steps: blocks permutation, rotation and inversion, negative and positive transformation, and color channel shuffling. Based on these steps, several methods have been proposed in literature. They can be classified as Color-PE and Grayscale-PE methods based on their input image representation. In Color-PE methods, an input color image is represented as a three-component image, whereas Grayscale-PE methods represent an input as pseudo-grayscale image by concatenating its color components along the vertical or horizontal direction. In Grayscale-PE methods, the channel-shuffling step is omitted. This section provides a summary of PE related work.

Kurihara et al. [17] proposed a block-based Color-PE method that performs the encryption steps on each color component by using the same key. Since, the input is a color image, larger block size is used to avoid block artifacts in a decoded image resulted by chroma subsampling of the JPEG algorithm. However, the use of the same key for each color component and larger block size result in a smaller number of blocks, which make the scheme vulnerable to jigsaw puzzle attack. To increase the number of blocks for better security, Imaizumi et al. [43] proposed to perform the first three steps of encryption independently in each color component. As a result, the scheme has a larger key space than that of [17]; however, processing each component individually results in the JPEG compatibility issues. For example, the method is only applicable with the JPEG lossless algorithm only when using RGB colorspace. Ahmad et al. [12,18] proposed a PE method to deal with the compatibility issue of [43]. In their proposed schemes, rotation, inversion and pixel values transformations are performed on each color component independently. The use of a same key for permutation step in each component allows the JPEG algorithm with YCbCr colorspace for better compression savings. The extended PE methods in [12,18,43] have better security than the PE method proposed in [17] as the keyspace is expanded and color distribution is altered significantly. However, the main limitation of extended PE methods is that they cannot exploit the JPEG chroma subsampling.

An alternative approach has been adopted by Chuman et al. [13] that allows the use of a smaller block size. The main idea is to represent an input color image as a pseudo grayscale image by concatenating the color components along the horizontal or vertical direction; therefore, belongs to Grayscale-PE methods. Such representation allows use of a smaller block size without any adverse effect on the decoded image quality and compression savings. In addition to smaller block size, lack of color information improves robustness against jigsaw puzzle solver attack. When chroma subsampling is desirable, Sirichotedumrong et al. [14] proposed to convert input image from RGB to YCbCr colorspace, perform down sampling on the color components and then concatenate them with the luminance component. For example, to combine them horizontally, the color components must be concatenated vertically and vice versa.

Compared to Color-PE methods and their extensions, Grayscale-PE methods provide better security as their use of smaller block size increases the number of blocks and pseudo grayscale representation efficiently disrupts the color information. However, for block-based PE schemes, there is an efficiency tradeoff between compression and encryption because of the block size choice. For example, a block size of no smaller than 16 × 16 and 8 × 8 should be used in Color-PE and Grayscale PE methods, respectively.