Hip fractures, mainly displaced femoral neck fractures, pose a significant diagnostic challenge due to their elusive nature on conventional radiographs. These fractures often evade detection in initial assessments, underscoring the critical necessity for refined classification systems. This paper presents a neural network model that can classify X-ray hip fractures accurately and focuses on the need for explanations to increase trust and interpretability. As a first contribution, we evaluate the suitability of several model-specific explanation methods for this concrete classification task. To enhance the limited explainability of these model-specific methods, we also present a surrogate model applying Case-Based Reasoning (CBR) to perform explanations by examples enriched with textual descriptions. This CBR surrogate uses a clustered case base with prototypical cases that include textual medical reports of the fractures. The last contribution of the paper is the integration of LLMs into the reuse stage of the CBR cycle to adapt and personalize these textual reports to the query image and target user.
Interpretability and trust have become a requirement for black box models applied to real-world tasks like diagnosis or decision-making processes. In the medical domain, interpretability ensures that medical practitioners can validate and explain the reasoning behind AI-driven diagnoses. This transparency not only enhances the credibility and acceptance of AI technologies within the medical community but also facilitates collaboration between human experts and machine learning systems.
Our research aims to enrich AI-driven diagnoses with diferent types of explanations. Firstly, we have explored the applicability of several model-specific explanation methods, such as Grad-CAM, that highlight crucial regions of an image that influence the network’s prediction based on the gradients of the last convolutional layer.
However, this approach has limited explainability, and therefore, we propose a surrogate system
based on CBR [
Section 2 describes related work on the hip fracture detection problem and its interpretability. Section 3 describes our fracture classification system, which uses a deep learning approach trained from a heterogeneous dataset of X-ray images. As it is a black-box system, we describe gradientbased explanations in section 4. In Section 5, we define a CBR system to generate natural language explanations without an extensive corpus of textual data for training. The system first retrieves a prototypical explanation that is extended using a generative LLM. Finally, Section 6 concludes the paper and discusses open lines of future work.
The classification of hip fractures is a critical aspect of orthopedic diagnosis, guiding treatment decisions
and patient care. With the advent of artificial intelligence (AI), there has been a surge in research aimed
at automating and improving the accuracy of hip fracture classification. In this section, we review
relevant literature, including the pioneering work of Krogue et al. [
Justin Krogue et al. [
The results of Krogue et al.’s research were highly promising, with the developed model achieving remarkable accuracy rates. The binary accuracy for detecting a fracture reached 93.7%, with a sensitivity of 93.2% and a specificity of 94.2%. Moreover, the multiclass classification accuracy was reported at 90.8%. Remarkably, the model’s performance was comparable to or even surpassed that of human experts, including fellowship-trained radiologists and orthopedists, under various conditions. As for explainability, point-based gradient explanations were applied to the model. It is noted that the model appears to pay attention to cortical outlines to make its classification, while the lucent fracture line appears to receive very little attention.
Furthermore, Krogue and his team demonstrated that the use of the deep learning model significantly improved human performance in hip fracture classification. When aided by the model’s predictions, resident physicians achieved classification accuracies approaching those of unaided fellowship-trained attending physicians.
The work of Tanzi et al. [
The work conducted by Twinprai et al. [
Previous successful hip fracture detection studies include the models of Cheng [
Regardless of the quality of diagnosis, trust and acceptance are vital to the adoption of medical AI systems, and these depend on the ability to explain the system’s decisions. Our focus lies in pushing the boundaries of application by integrating a CBR model to generate diagnostic insights from radiographic images.
The work conducted by Kim et al. [
Our approach proposes ANN-CBR Twins [
Our hip fracture classification system uses a deep learning approach trained from a heterogeneous dataset of X-ray images. Next, we describe this dataset, its preprocessing, and the resulting classification model.
Our starting point is the AO/OTA1 classification for bone fractures [
Our primary goal was to train a classification model for hip fractures. To do this, we needed to
acquire correctly labeled data in the form of hip X-rays. Thanks to the Virgen de la Victoria Hospital
and Müller Foundation Spain, we have collected this data from three diferent sources:
• The YOLO Roboflow project available online at [
No personal data was collected, all the data was completely anonymous when received.
Since our collection of X-rays came from three diferent sources, we had to deal with the problem of obtaining a uniform and homogeneous dataset with which to train our model. To achieve this, we performed various methods of preprocessing on the images. The data from the Müller foundation was the most challenging to preprocess as a portion of the data (n=62) had an inverted color palette, meaning the image displayed a black bone on a white background. In contrast, the rest of our images had a black background with white bones. Furthermore, most of these images came with two X-rays: one showing the bone fracture and another showing the prosthesis or implant used to treat the patient. This is shown in Figure 2.
An additional problem with the collected images is the cropping of the femur area. As Figure 2 (right) illustrates, most of them were general X-ray images where the femur area was only a small portion of the whole image. Therefore, we required a method to crop the hip area.
For this task, we obtained the femur crops using a custom YOLOv5 model trained on our complete dataset. Ultimately, we implemented Python scripts to flip right femurs into left femurs, unify color scale into gray, and resize images to 224x224 pixels, which are the dimensions required for the input channels of our residual neural network.
A diagram representing the obtention of the whole dataset can be found in Figure 3. As shown in the diagram, we finally obtained 891 images corresponding to unbroken hips, 212 to femoral neck fractures, and 294 to hips presenting a fracture in the trochanteric region.
The collection of such a challenging dataset led us to take the following additional preprocessing steps: • Given that the data is generally only labeled as type A, type B, or type C, it was not possible for us to train a classifier for the given subtypes of A, B, or C. Even if we had the complete labels, we 1AO/Orthopaedic Trauma Association. AO stands for Arbeitsgemeinschaft für Osteosynthesefragen or the Association of the Study of Internal Fixation (ASIF), established in 1958 by a group of Swiss general and orthopedic surgeons to strive to transform the contemporary treatment of fractures in Switzerland. Color inversion Manual crop
n = 62
Class 0 No fracture n = 891
Flip right femurs Homogenize grayscale
Resize to 224x224
Class 2 Trochanteric fracture n = 294 did not have enough data to implement a model able to predict the subtypes, so we decided to focus on just the general type, forgetting the subtypes. Moreover, type C fractures are unusual, which results in a three-class model, predicting no fracture, type A fractures, or type B fractures. Working beyond this problem has no easy solution, as it involves experts manually labeling each X-ray into more complicated types and gathering more data. • Looking at the data, it is obvious we have an imbalanced class problem: for types 1 and 2, we have a similar amount of data, but for type 0 (unbroken), we have about three times as much data. To solve this, we used the Albumentations library, which has been designed for image augmentations, primarily focused on deep learning tasks. It ofers a wide range of transformations from where we implemented uniform angle rotations between 0 and 15 degrees and random gamma, adjusting the global brightness of the image. These transformations were applied with a probability of 50% and 10%, respectively. We perform three augmentations for classes 1 and 2, and this way, we finally obtain a balanced dataset with about 2600 images ready to train our model with 891 images for class 0 and about 800-900 for each of classes 1 and 2.
We performed tests on diferent architectures of neural networks, finally opting for a Residual Network with 18 convolutional layers (Resnet18). This model has an 18-layer structure that incorporates residual blocks with skip connections, addressing the vanishing gradient problem and enabling efective training of deeper networks. Pre-trained weights available for ResNet18 further enhance its utility through transfer learning, allowing fine-tuning on medical datasets with reduced labeled data requirements. By automatically extracting relevant features from medical images, ResNet18 proves adept at capturing bone structures, edges, and textures crucial for fracture identification. We have trained this model for 10 epochs, implementing cross-validation to avoid over-fitting. We split the dataset into 80% to train and 20% to test, and we obtained satisfactory results as reported in Table 1 and the confusion matrix shown in Figure 4. Although the score of the model is very high, we have ruled out the posibility of over-fitting for two main reasons: firstly, the results on the test set are impressively satisfactory, and secondly, helped by experts of the University Hospital Virgen de la Victoria, we have noticed that the explanations generated by Grad-CAM, and detailed in the following section, highlight regions close to the location of the fracture.
Weighted avg
As a classifier, the model ofers no means of interpretability, a fundamental cornerstone in medical systems. One common approach for explainability when working with image data is applying Gradientweighted Class Activation Mapping (Grad-CAM) or Integrated Gradients, and it was our first focus.
Grad-CAM analyzes gradients in the last convolutional layer to decode the importance of each
feature map for a specific class. It generates a heat map, highlighting crucial regions of an image that
influence the network’s prediction, providing interpretability without compromising accuracy. By
visualizing how models arrive at predictions, we gain insights, debug models, and enhance performance.
Several heat-map explanation methods from the libraries Xplique 2 and pytorch_grad_cam 3 have been
implemented and compared. The ones from Xplique were discarded because, under the criteria of
experts at Virgen de la Victoria Hospital, the results were not satisfactory, as those explanations did not
seem to highlight the regions where fractures were visible. To choose between GradCam, Gradcam++,
EigenCam, and Ablation, the ROAD (RemOve And Debias) score was implemented. Fundamentally,
this metric gives a score to a heatmap by obscuring the more important regions and performing a
new prediction to understand how hiding such regions afects the output. This way, if hiding the red
zone of a heatmap means a change in the model’s prediction, we can afirm this region is relevant to
the model so that this explanation will have a high score. More information about this metric can be
found in Rong et al. [
Here, LLMs appear as a direct and efective solution. However, current state-of-the-art generative models, such as GPTs, are of no use in this case since they have not been trained with medical images, and fine-tuning requires a large corpus of textual explanations associated with X-ray images. As the manual collection of this corpus is unviable, we proposed a hybrid solution using CBR, as presented next.
Our research faces a pivotal question here: besides gradient-based explanations, we have so far, can we efectively generate natural language explanations for this AI-driven fracture classification without an extensive corpus of textual data for training?
As we only have a limited number of textual explanations associated with prototypical fractures, CBR raises as a suitable approach to, given a new X-ray image, retrieve the most similar prototypical case and reuse its textual explanation. The adoption of the CBR paradigm has the additional benefit of enabling the use of the most similar images as explanation examples. This way, our explanation methodology is three-fold: attribution-based (Grad-CAM), textual (LLMs), and explanation-by-example. All these explanation modalities are presented to the user through the graphical interface presented in Figure 7.
Therefore, we propose a twin surrogate architecture to explain the classification model. Figure 6 depicts the two parallel pipelines. The first (top) pipeline performs the classification process using the Resnet model after preprocessing the images with YOLOv5 (see Section 3), and provides a visual explanation using Grad-CAM (see Section 4). In the second pipeline, we use a CBR system to retrieve prototypical explanations followed by an LLM that adapts the resulting text to generate textual explanations. These processes are described in the next sections.
The case base has been elicited by an expert who provided highly representative images and the corresponding medical reports, which we will also call prototypical cases. These will be the seed cases of our case base, defined as = ⟨, , ⟩, where is the X-ray image, the classification given by the ResNet18 model, and is the associated textual report.
Initially, we have about ten seed cases that have been included with the help of experts: one image of an unfractured femur, three images of subtypes of femoral neck fractures, and two images of subtypes of trochanteric fractures with their corresponding texts. Each of these texts is a definition for the given subtype of fracture. We also included three general explanations for femoral neck fractures and two for trochanteric fractures. Figure 8 shows an example of a prototypical case. In this text, the part corresponding to a general description of a femoral neck fracture is highlighted in blue, while below, in gray, appears the text explaining the subtype of fracture that, according to the expert, is recognized in the prototypical X-ray shown on the right.
Due to the limited number of seed cases, we propose a retrieval strategy using a clustered organization of the case base. Having limited textual explanations, we aim to spread such texts to a wider range of images, expanding our base cases. We implement an algorithm to cluster images within the same type of fracture, keeping at least one prototypical case in each resulting cluster.
The idea is to group around the prototypical cases, which would serve as "centroids", providing their explanation to the associated cases. Initially, our idea was to implement a KMeans algorithm, but unfortunately, this would move the centroids, meaning that prototypical cases could end up in a diferent cluster and some clusters may not get a prototypical case at all. Femoral neck. This area is located at the top of the femur, just below the spherical part (head of the femur) of the ball-andsocket joint. Fractures of the femoral neck are divided into basicervical or transcervical depending on the distance to the femoral head, basicervical fractures occur at the junction of the head and the neck, and transcervical fractures occur in the neck itself. Since the blood supply to the femoral head passes through the neck, this type of fracture can cause a complication because the break often blocks the passage of blood to the femoral head, leading to avascular necrosis, which is equivalent to a heart attack but in bone. These fractures can occur in young people in high-energy accidents but are common among older individuals who fall from their height due to osteoporosis.
A transcervical femoral neck fracture is another specific type of fracture that occurs in the femur bone, in the region of the femoral neck. In this case, however, the fracture traverses the femoral neck from side to side, completely dividing it. It can have serious consequences due to the disruption of blood flow to the femoral head.
The clustering process is as follows. Wanting to respect the label restriction, every image of a given
label is compared only with the prototypical cases of such label, associating it with the one to which it
is most similar, based on the Structural Similarity Index Measure (SSIM) metric [
In this stage, we can retrieve the prototype case for the cluster where the query image is classified. Firstly, the query goes through our Neural Network classifier, providing a predicted label for it. Secondly, once the label is obtained, the image is compared to all the cases in that same label. As a result, we retrieve the most similar image, and, since it already belongs to a cluster, we can provide its explanation for our query.
In Figure 9, the diferent clusters generated for each of the classes are shown, along with the image representing the centroid of each one. Each cluster is built around a centroid that represents a certain subtype of a fracture, blue being subtypes of femoral neck fractures and red ones for fractures in the trochanteric region. As mentioned in this section, the idea is that images sharing the subtype within the same class will exhibit some similarity among them and, therefore, should belong to the same cluster or be close to each other. This way, we manage to propagate the explanations we have to new images.
As we have mentioned, for comparing images, we use the metric known as the Structural Similarity Index Measure (SSIM), which is measured between -1 (completely diferent images) and 1 (identical images). This measure is particularly interesting for our domain because, in addition to considering parameters such as luminosity, contrast, or brightness of the images to determine their similarity, it also takes into account the structure of these images, which is perhaps the most crucial aspect when comparing X-ray images. Class 0 Class 1
Class 2 0.246
Average SSIM per cluster
Figure 10 reports an evaluation of the quality of the resulting clusters, measured through the average SSIM similarities within each cluster. As we can observe, all the clusters have an equivalent similarity, denoting a balanced distribution of the cases.
Using the clustering approach has the additional advantage that the image corresponding to the prototypical case of the cluster can be presented to the user as an explanation example. This way, our explanation approach not only uses the heat maps provided by Grad-CAM and the textual explanations generated by the LLMs but also provides an explanation case to ease the understanding of the fracture.
Recall that a simple search by SSIM will not provide an accurate classification of a given subtype of fracture. SSIM retrieves images that are structurally similar, but it may not recognize the underlying diferences between fractures. Since the subtype explanations are generated based on an SSIM search, they may not be reliable or precise while the general ones will always be, as the images are taken from the same label.
Finally, Figure 11, shows an example of the retrieval stage: the most similar image to a given one is displayed, along with the cluster where this latter image is located. The classifier’s prediction is used to determine among which images to search for the most similar one since, as mentioned earlier, it is only compared with images within the same class. 5.3. Reuse The reuse stage is coupled to an LLM to adapt the textual report of the retrieved case. This is useful to prevent the system from being too repetitive, especially in the beginning when there are few cases. Note that the use of generative models to provide textual explanations has the risk that it may generate some invalid explanations, which is why the review phase of the CBR system is crucial in our proposal.
To implement this stage, we have tested Mistral and GPT-4 through the HuggingFace and OpenAI APIs, respectively. The prompts include the textual report obtained by the retrieval stage and various "knowledge levels" of the user, with which we intend to generate outputs adapted to diferent types of users (medical experts or patients/students).
Table 2 shows two diferent versions generated by GPT-4 of a case base text created by an expert to explain a basicervical femoral neck fracture. In the second column, the original text is presented, and in the fourth column, text versions for students and experts, respectively, generated by GPT-4 are shown. This approach helps us avoid repetitive explanations, as GPT-4 generates diferent texts each time. Furthermore, we can adapt the original text depending on the user, whether they are a student or an expert, thereby enhancing the usability of the system.
Expertise
Specific diagnosis
Prompt
Generated diagnosis
A basicervical femoral neck fracture is a specific type of fracture that afects the base of the neck of the femur, at the junction with the trochanteric mass. It is a two-part fracture at the base of the femoral neck that was medial to the intertrochanteric line, exiting above the lesser trochanter but was more lateral than a classic transcervical fracture.
medical explanation: «Specific diagnosis», generate a text of
length that is suitable for a person with Student knowledge.
medical explanation: «Specific diagnosis», generate a text of
length that is suitable for a person with Expert knowledge.
A basicervical femoral neck fracture is a specific type of fracture that afects the base of the neck of the femur, at the junction with the trochanteric mass. It is a two-part fracture at the base of the femoral neck that was medial to the intertrochanteric line, exiting above the lesser trochanter but was more lateral than a classic transcervical fracture. A basicervical femoral neck fracture is a break in the bone that occurs where the thigh bone’s neck and its wider part at the top meet. Unlike a common neck fracture that happens further inside, this type takes place closer to the side, above a bony prominence but not as far to the side as some other types.
In our approach, the revision phase is not merely a formality but a critical mechanism for refining and enhancing the performance of our CBR system. With an initial case base comprising only a handful of seed cases and a reuse stage reliant on LLMs, which may occasionally produce inaccurate textual descriptions, the revision stage emerges as a pivotal component of our methodology.
Our system has been designed with a user-centric approach, empowering experts and practitioners to engage in the learning process actively. One of the key features of our system is its capability to solicit alternative textual descriptions directly from users during the revision process. This functionality enables users to provide nuanced insights, corrections, or additional context that our system may have overlooked or inaccurately represented.
By incorporating this interactive feedback loop, we not only facilitate user involvement but also foster a process of continuous improvement. Every correction, refinement, or enhancement suggested by users serves as a valuable input for the system, contributing to its iterative learning process. Over time, as the system accumulates diverse user-contributed insights, it evolves into a more robust and adaptive tool for diagnosing hip fractures from X-ray images.
This way, the review stage serves as a cornerstone of our CBR approach, facilitating ongoing refinement and enhancement through active user engagement. By embracing user feedback and leveraging the collective intelligence of the medical community, we strive to create a system that not only meets the needs of its users but also evolves and adapts to deliver increasingly accurate and reliable diagnostic capabilities.
Due to the low number of seed cases in the current state of the case base, we allow the direct addition of new cases to the case base and we retain each revised case as a new case for future iterations. We plan to extend the research to evaluate how these new cases can afect the system behavior and other approaches for the retain phase. It is important to note that the inclusion of new cases potentially implies recalculating the clusters that organize the case base.
Hip fractures on conventional radiographs pose a significant diagnostic challenge. This paper first presents a neural network model that can classify X-ray hip fractures accurately and focuses on the need for explanations to increase the trust and interpretability of such a black-box classifier. We propose a twin surrogate architecture to elucidate the classification model. This architecture comprises two parallel pipelines. The first pipeline conducts classification and provides a visual explanation using Grad-CAM. The second pipeline consists of a CBR system to retrieve prototypical explanations, followed by an LLM that adapts the resulting text to generate textual explanations.
We use a seed case base of prototypical X-ray images for hip fractures manually annotated with textual explanations by an expert. The CBR system retrieves a prototypical case and uses an LLM to generate extended textual explanations. This way, our explanation approach is three-fold: attributionbased (Grad-CAM), textual (LLMs), and explanation-by-example. While integrated gradients serve as a valuable tool for highlighting important regions in X-ray images, textual explanations ofer a more comprehensive and personalized approach to understanding. By providing contextual insights, the interpretation process is more efective. Integrating both approaches can enhance the overall interpretation process, ensuring accurate diagnosis and informed decision-making in clinical practice.
As future work, as the case base grows, scaling up this CBR system poses several challenges and considerations. First of all, storing and managing a vast number of diagnostic texts requires robust infrastructure and eficient data-handling techniques. Therefore, it is essential to develop a method for selecting which textual explanations provided by experts are crucial for refining the CBR system or for combining similar cases. This selection process could involve prioritizing explanations based on their relevance to common diagnostic scenarios, their impact on improving model performance, or their alignment with evolving clinical practices. The use of diferent types of prompts and a deeper study on how to personalize the textual explanation to the query are both a challenge and an opportunity for the impact of synergies of CBR and LLMs in the eXplainable AI (XAI) landscape.
Furthermore, with an expanding dataset of diagnostic texts, another aspect to address is the dificulty of reorganizing clusters maintaining an eficient and meaningful clustering structure as new cases are added. For this reason, it would be necessary to decide when the image groups are recalculated to generate better results as the system is used.
Another challenge related to the dataset growth lies in the process of searching for similar images. Given an image, this CBR system has to compare it with all the others with the same label. Too much time could be spent on this task, so a way of speeding up this process becomes necessary.
On the other hand, collaboration between domain experts and programmers will be crucial in addressing errors generated by LLMs, which present a significant challenge. Despite their advanced capabilities, LLMs are not immune to inaccuracies or misinterpretations, so correcting these errors is essential for ensuring the reliability and accuracy of generated textual explanations. To make the most of experts’ feedback and refine stored diagnostics, it’s crucial to establish a method promptly once this CBR system gains widespread usage. In line with this, it’s essential to share explanations with experts via meetings and surveys. This not only confirms the accuracy of the explanations but also ensures the efectiveness of our approach for patients. This step of the process is designed specifically for medical expert users capable of precisely drafting medical reports.
This work is supported by the PERXAI project PID2020-114596RB-C21 funded by MCIN/AEI/10.13039/501100011033 and the Complutense University of Madrid (Group 921330). Special thanks to the Orthopedic Surgery and Traumatology department at Hospital Universitario Virgen de la Victoria and to the Maurice E. Müller Foundation.