To concurrently and adaptively determine the importance of key regions and various channels in CXR images, this paper combines the channel attention mechanism with the designed enhanced spatial attention mechanism in the residual block to form a multi-attention mechanism to achieve this goal. Fig. 4 shows the structure of the residual block (RCESA) with the multi-attention mechanism.

In Fig. 4, incorporating residual blocks into the network allows for image feature extraction and helps tackle the issue of network degradation. Taking the intermediate feature vector M_in extracted by the neural network as input, the weight of important areas is first enhanced through the enhanced spatial attention mechanism, and then its output serves as the basis for the channel attention mechanism operation to improve the accuracy of feature extraction. Specifically, within each residual block, different-sized convolutional layers are first used to extract multi-scale features to obtain the multi-scale feature vector M_F. Then, the extracted features are subjected to max-pooling operation (used for further downsampling of the feature map after partial feature extraction to reduce network computational load and enhance the network model’s abstraction ability for important features of CXR images). Next, features are further extracted through a 3 × 3 convolutional layer, and bilinear interpolation is used for upsampling to match the feature map size with that of the input. The upsampled operation is then combined with the preceding feature map to generate the feature fusion vector M_A. Feature fusion is performed through 1 × 1 convolutional operation, allowing the network to more effectively utilize features of varying scales. Ultimately, the feature vector after fusion is normalized to the [0, 1] range with the sigmoid activation function to obtain the weight matrix m, which is subsequently multiplied by the input feature M_in to generate the weighted feature map as the output M_out of enhanced spatial attention operation, achieving enhanced spatial attention to enable the network model to utilize crucial regional information in input CXR images more effectively. This series of operations is represented by Eq. (5): (5)Mout=x×Sigmoid(Conv(Conv(Interpolation(MaxP(Conv(Conv(x))))+Conv(Conv(x))))).

Subsequently, the feature vector M_out processed by the enhanced spatial attention mechanism is used as the input of the channel attention mechanism. Initially, the convolutional feature vector M_G is obtained through two convolutional layers. This is followed by max-pooling and average-pooling to generate two vectorized representations. These vectors are then inputted into a shared multi-layer perceptron (MLP). The representations generated by the MLP are summed to form an attention vector, which is then multiplied with M_G to produce the feature map M_H with channel attention. This series of operations is represented by Eq. (6): (6)MH=σ(φ(ω1(MG))⊕φ(ω2(MG)))⊗MG,where σ is the sigmoid function, ϕ represents a trainable transformation within the MLP, ω₁ and ω₂ respectively represent max-pooling and average-pooling operations. After this step, the shortcut connection defined by Eq. (7) is applied to obtain the output of the RCESA. (7)Me=Mout⊕MH.

Each RCESA’s MLP comprises two convolutional layers with a kernel size of 1 × 1, referred to as Layer-1 and Layer-2. The detailed settings of the MLP are shown in Table 1.

To understand the complex relationships among the input samples, triplet constraints serve as the loss function guiding the training of deep networks, improving the distinguishing ability of the obtained deep feature hash codes. The post-training CXR images are represented as {X₁, …, X_i, …, X_M}, and the associated class labels are represented as {L₁, …, L_i, …, L_M}, where L_i∈{1, …, c}, and c represents the number of categories. Fig. 5 shows an example of triplet training.

As shown in Fig. 5, the triplet loss is designed to increase similarity among samples from the same class while decreasing similarity between samples from different classes [38]. Mathematically, given a triplet unit {q, p, n}, where q represents the query sample, p represents the positive sample, and n represents the negative sample. The goal of triplet loss is to reduce the following items: (8)d(q,p)−m≤d(q,n),where d(q, p) is the distance between the query and the positive sample, d(q, n) is the distance between the query and the negative sample, and m represents a preset positive margin value.

In this study, given an image Xiq, triplet {Xiq,Xip,Xin} is generated through random selection. To attain a more distinctive representation for effective CXR image retrieval, this paper simultaneously applies triplet constraints to both deep features and hash codes, as outlined in Eq. (9): (9)ℒT=∑i=1T[max(0,m−d(hiq,hip)+d(hiq,hin))+max(0,m−d(fiq,fip)+d(fiq,fin))],where T indicates a total number of triplet units across the training images, hiq, hip and hin denote the hash codes of the query image Xiq, positive sample image Xip, and negative sample image Xin in the i-th triplet unit, respectively. fiq, fip and fin represent the deep features of the query image Xiq, positive sample image Xip, and negative sample image Xin in the i-th triplet unit individually, m denotes the margin value, and d(⋅) represents the Euclidean distance.

The specific learning algorithm for MATDH is shown in Algorithm 1.

Mean Average Precision (mAP) is commonly used to evaluate object detection models, considering the average precision (AP) across different categories; a higher mAP indicates more similar images at the top of the retrieved list. Additionally, P@H ≤ 2 is a key measure in image retrieval, emphasizing the use of shorter hash codes to capture similarity within the feature space. This paper calculates the average precision for instances where the Hamming distance between the query and database images is no more than 2. To mitigate bias from model initialization, five independent experiments were conducted.

To validate the CXR image retrieval performance of the proposed method, a comparison was conducted between our MATDH method and six other deep hashing methods, namely TCDH [8], SWTH [12], ATH [22], DPN [40], ASH [41], and DBDH [42]. For fairness, the comparison was conducted using the same data and parameter settings as MATDH. Table 2 shows mAP values for different methods across various hash code lengths.

As can be seen from Table 2, it can be observed that MATDH outperforms the compared existing deep hashing methods. For example, with a 48-bit hash code length, MATDH (0.9117) enhances the mAP by approximately 4.2% over the top-performing deep learning technique, TCDH (0.8691). This is because the features extracted from CXR images after image enhancement under the dual attention mechanism are more accurate, and guided by the triplet loss function, the network can extract more discriminative features. Furthermore, after performing a t-test distribution calculation comparing the mAP of the MATDH method with that of six other methods, it was found that there is no significant difference between MATDH and ATH, while significant differences exist with the other five methods. This indicates that the MATDH method has statistical significance, providing important evidence for selecting and optimizing medical image retrieval methods. Fig. 8 shows the curve of P@H ≤ 2 for MATDH compared to six other methods.

As can be seen from Fig. 8, P@H ≤ 2 curves further illustrate the method’s effectiveness within a Hamming radius of 2. The accuracy curve in Fig. 8 shows that the proposed MATDH method consistently outperforms other advanced methods across hash code lengths from 16 to 64 bits. At a hash code length of 48 bits, the accuracy is approximately 20% higher than that of the DPN method. This suggests that the MATDH method is capable of retrieving a greater number of images from the same category when the Hamming radius is set to 2. Moreover, in comparison to other methods, the precision attained by MATDH consistently exceeds 0.85 as the number of hash bits increases. This highlights MATDH’s effectiveness in CXR image retrieval and its resilience to Hamming sorting.

Additionally, with a 48-bit hash code length, MATDH achieves the highest mAP values. This suggests that for CXR images, using 48-bit hash codes with MATDH can better preserve the main features of the images. Therefore, the MATDH method in this paper adopts a 48-bit deep hash code.

Besides relying on mAP, in medical image retrieval, accuracy for items positioned highest in the ranking is often a key priority for users. Therefore, this paper also uses the recall rate in the top-K retrieval results (Recall@K) and the precision of the top-N retrieval results (P@N) to evaluate the effectiveness of CXR image retrieval methods [8].

Fig. 9 shows the recall rate curves of top-k images across various techniques using a consistent 48-bit hash code length, and Table 3 presents the comparison results of P@1, P@5, and P@10 obtained through various methods using 48-bit deep hash codes in the COVIDx dataset. The XMIR method [43], which is not a deep hash method, is not concerned with the issue of hash code length.

As shown in Fig. 9, the recall rate of MATDH remains consistently higher than that of other methods as the number of returned CXR images increases from 0 to 200. For example, when 180 images are returned, the recall rate is approximately 30% higher than that of the DBDH method. This indicates that MATDH retrieves a higher number of relevant images, consistent with the comparative analysis results.

As can be seen from Table 3, it is evident that when the limit on returned images is set, MATDH outperforms other methods. This is because, under the effect of the multi-attention mechanism, with the iterative training of deep neural networks, the constructed hash codes can more accurately represent images, thereby improving the accuracy of CXR image retrieval.

Although this study mainly uses the COVIDx dataset for experiments, the proposed MATDH method has the potential for clinical application. It can serve as an auxiliary diagnostic tool, helping doctors make faster and more accurate judgments on COVID-19 infections. The model can be integrated into hospital imaging systems for real-time radiological image analysis, improving early detection accuracy. Additionally, the proposed MATDH method can assist doctors in quickly and accurately retrieving medical imaging data, reducing the workload of data queries during treatment, especially when medical resources are limited.

To objectively evaluate the network model and framework of this paper’s method, disintegration experiments are conducted through the removal or substitution of modules in the network architecture. For the network structure, the basic structure of ResNet18 in the UNet network used by the proposed method is replaced with AlexNet [44] and ResNet34 [45], and these two variants are respectively denoted as MATDH-AlexNet and MATDH-ResNet34. Furthermore, before extracting features from the deep network, the approach described in this paper does not enhance the original images, which are denoted as MATDH-NoEnhanced, and experiments are conducted using 48-bit deep hashing codes. The detailed data can be found in Table 4, where the top entries are emphasized in bold.

As can be seen from Table 4, MATDH enhances the retrieval performance of input CXR images by utilizing image enhancement, and it outperforms variant networks with different frameworks.

To examine how the dual attention mechanism within the RCESAs module influences the outcomes of CXR image retrieval experiments, this section modified the attention mechanisms used in the RCESAs module of the MATDH framework. One of the variants utilizes only the channel attention mechanism, referred to as MATDH-CA, while the other variant utilizes only the enhanced spatial attention mechanism, referred to as MATDH-ESA. Fig. 11 shows the comparison between MATDH and its variants that utilize various attention mechanisms for CXR image recovery from the COVIDx dataset, using a 48-bit hash code length, with the top outcomes emphasized in bold.

According to the results in Fig. 11, MATDH with a dual-attention mechanism outperforms the standalone CA and ESA mechanisms in retrieval performance. The specific contributions to this performance improvement can be summarized as follows:

1) Feature enhancement: The dual-attention mechanism effectively captures and enhances the correlation between useful feature channels, enabling the model to better identify features related to lesions in CXR images, thereby improving retrieval accuracy.

2) Redundant information suppression: This mechanism effectively reduces the influence of unnecessary data and noisy channels in CXR images, allowing the network to focus more on key information, thus enhancing the clarity of feature representation.

3) Dynamic attention adjustment: The dual-attention mechanism dynamically adjusts the focus of the deep neural network on different regions of CXR images, enabling the model to better capture important spatial features. This flexibility allows MATDH to maintain high retrieval performance across a diverse set of input images.

4) Combined effect: By integrating the above factors, the dual-attention mechanism significantly enhances the overall performance of MATDH, demonstrating its effectiveness and superiority in medical image retrieval tasks.

In cryptography, information entropy indicates the degree of unpredictability in image data. For a perfectly random image, it is essential that all grayscale values have an equal probability of occurrence. The formula for calculating information entropy is shown in Eq. (14):

(14)H=−∑i=0Lp(i)log2⁡p(i),where L represents the total number of grayscale levels in the image, while p(i) denotes the probability of occurrence for grayscale value i. For an image with p(i) = 256, when each pixel value in the image has an equal probability of occurrence, the entropy of the image can reach its maximum value of 8. A higher entropy indicates greater randomness in the image, leading to better encryption performance. Table 5 shows the compares the entropy between the original and encrypted images, both with dimensions of 512 × 512.

As can be seen from Table 5, the average entropy of the encrypted images generated by the encryption method presented in this paper is 7.9993, approaching the optimal value of 8. This suggests that the encryption technique presented here achieves strong performance, resulting in encrypted images that show significant randomness.

As the encryption method described in this study is a secure lightweight encryption algorithm, this section compares and analyzes it with existing schemes based on three evaluation indicators reflecting the lightweight and security of encryption algorithms: key space size, encryption speed, and information entropy. Table 6 shows the performance comparing results between the encryption method introduced here and various medical image encryption schemes [46–48].

From Table 6, it is evident that the encryption method in this paper shares the same key space as in Castro et al. [46], but is significantly smaller than that in Abdelfatah et al. [47] and Inam et al. [48]. This suggests that the proposed method is more resource-efficient, making it suitable for lightweight encryption of medical images. Although its key space is smaller, it still provides sufficient security (greater than 2¹⁰⁰), enough to resist exhaustive attacks [48]. The encryption speed is higher than in Castro et al. [46] and Abdelfatah et al. [47] and slightly lower than in Inam et al. [48], due to frequency domain operations. These operations, while slowing down the process, offer better attack resistance. Additionally, the proposed method has higher entropy, indicating strong randomness and resistance to entropy attacks, making it a secure lightweight encryption algorithm for medical images compared to other schemes [46–48].

Building on the insights from the previous two sections, this part demonstrates the advantages of the encryption algorithm in resisting various types of attacks.

First, as discussed in Section 4.6.1, it is understood that the ciphertext generated by this algorithm exhibits high entropy, meaning that it is statistically close to a random distribution. This randomness complicates the process for attackers trying to obtain useful information through frequency analysis, as well as through statistical attacks such as known-plaintext or chosen-plaintext attacks. The existence of high-entropy ciphertext significantly enhances the algorithm’s defense against these forms of attacks.

Secondly, the analysis in Section 4.6.2 shows that the key space of this encryption algorithm is 2²⁵⁶, and such an extensive key range renders brute-force attacks virtually impossible. Even if attackers possess powerful computational resources, attempting to exhaust all possible keys within a reasonable time frame is still unlikely, providing additional security for the algorithm.

In summary, based on the aforementioned points, this encryption algorithm can be considered to possess strong security, effectively resisting brute-force attacks, statistical attacks, known-plaintext attacks, chosen-plaintext attacks, and chosen-ciphertext attacks, among other common attack methods. The high-entropy characteristic and the vast key space complement each other, ensuring the reliability and robustness of the encryption algorithm under various threats.