Abstract

Background:

Skin cancer is the most common cancer in the United States. Current estimates are that one in five Americans will develop skin cancer in their lifetime. A skin cancer diagnosis is challenging for dermatologists requiring a biopsy from the lesion and histopathological examinations. In this article, we used the HAM10000 dataset to develop a web application that classifies skin cancer lesions.

Methods:

This article presents a methodological approach that utilizes dermoscopy images from the HAM10000 dataset, a collection of 10015 dermatoscopic images collected over 20 years from two different sites, to improve the diagnosis of pigmented skin lesions. The study design involves image pre-processing, which includes labelling, resizing, and data augmentation techniques to increase the instances of the dataset. Transfer learning, a machine learning technique, was used to create a model architecture that includes EfficientNET-B1, a variant of the baseline model EfficientNET-B0, with a global average pooling 2D layer and a softmax layer with 7 nodes added on top. The results of the study offer a promising method for dermatologists to improve their diagnosis of pigmented skin lesions.

Results:

The model performs best in detecting melanocytic nevi lesions with an F1 score of 0.93. The F1 score for Actinic Keratosis, Basal Cell Carcinoma, Benign Keratosis, Dermatofibroma, Melanoma, and Vascular lesions was consecutively 0.63, 0.72, 0.70, 0.54, 0.58, and 0.80.

Conclusion:

We classified seven distinct skin lesions in the HAM10000 dataset with an EfficientNet model reaching an accuracy of 84.3%, which provides a promising outlook for further development of more accurate models.

Keywords:

skin cancer; pigmented skin lesions; machine learning; transfer learning; internet of medical things (IoMT).

Methods

Study Population

Dermatoscopy, also known as dermoscopy, epiluminescence microscopy, or skin surface microscopy, is a non-invasive, in-vivo method traditionally applied to evaluate suspicious skin lesions [20]. This well-used method improves the diagnosis of benign and malignant pigmented skin lesions compared to examination with the unaided eye [21]. The HAM10000 dataset is a set of 10015 dermatoscopic images collected over 20 years from two different sites, the Department of Dermatology at the Medical University of Vienna, Austria, and the skin cancer practice of Cliff Rosendahl in Queensland, Australia [14]. All data records of the HAM10000 dataset are deposited at the Harvard Dataverse. These Images and metadata are also accessible at the public ISIC archive through the archive gallery and standardized API calls (https://isic-archive.com/api/v1) [14]. The majority of lesions have been confirmed by pathology. At the same time, the ground truth for the rest of the cases was either follow-up, expert consensus, or confirmation by in-vivo confocal microscopy [14].

Image pre‑processing:

The data was obtained from Kaggle, available via a CC0: Public Domain License. It is appropriately anonymized and does not contain any identifiable features of the participants. As the dataset images were not labelled and were out of order, each image was first labelled using the dataset's metadata by transferring them into its respective folder. Then it was randomly split into a training set containing 8015 (80%) cases and a test set containing 2000 (20%) images.
images were resized to (240, 240, 3) tensors, as the EfficientNetB1 architecture, which was used to build the model, has the optimum performance with this size [22]. The well-known bilinear interpolation resampling technique was used in image processing to resize the images.
In order to artificially increase the instances, a data augmentation technique is used to generate new sample images. This technique consisted of a random width shift from -20% to +20% of image width and a random height shift from -20% to +20% of image height, and a random max 0.2-degree shear angle in a counter-clockwise direction to rectify the perception angle and also random horizontal and vertical flip was applied. Empty pixels were then filled by the nearest pixel. To better understand this procedure, we randomly applied data augmentation to 25 images 3 times, and the results are shown in Fig 1. Also, S1 File in supplementary materials illustrates 600 frames of random augmentations for 9 lesions with a frame rate of 10 FPS.

Model architecture for Transfer Learning:

It is generally not a good idea to train a very large Deep Neural Network from scratch as training such large models with at least 200 to 300 hidden layers requires a massive amount of resources and time not everyone has, instead using existing pre-trained models that accomplishes a similar task is a much reasonable idea. Transfer learning is a research problem in machine learning that focuses on storing knowledge in the process of solving one problem and applying the gained accumulated knowledge to accelerate the learning in new different, but related tasks [23, 24]. Transfer learning nets are trained on large datasets, and the model parameters of each layer could be manually set to be frozen so that they will not change during retraining.
The Efficient nets are a family of neural networks with the baseline model constructed with the Neural Architecture Search technique. The EfficientNET-B1, a variant of the baseline model EfficientNET-B0 which is created through compound scaling, is the backbone of our model. We deleted the top layer of EfficientNET-B1, then a Global average pooling 2D layer and a softmax layer with 7 nodes added on top. The model architecture is shown in Fig 2. As the EfficientNET-B1 has 340 layers, it is not possible to show all layers, so an expanded version of the architecture is available in supplementary materials.

Feature extraction and fine-tuning:

The top layer of the models has nodes equal to several classes and a softmax activation function which is used as the last activation function of a neural network to normalize the output of a network to a probability distribution over predicted output classes. In feature extraction transfer learning, this top layer gets replaced by a new softmax-activated layer with nodes equal to the new problem (7 nodes in this particular image classification problem). Then all other pre-trained model layers are set frozen so that only this new layer's parameters get trained to adjust the outputs to be more suited to a new problem.
At first, we freeze all the layers to train our new layers' parameters for feature extraction; then, the model is trained for 15 epochs. The learning rate was set at 0.001 primarily, but after 12 epochs, as the learning curve reached a Plateau, the learning rate reduced to 0.0001 for the remaining epochs. After feature extraction, the top 50 layers of EfficientNET-B1 were unfrozen to fine-tune the EfficientNET-B1 itself for our specific image classification problem. The learning curve is illustrated in Fig 3. The model was fine-tuned for 4 more epochs starting with the learning rate of 1.0e-4 and decreasing to 10% for each epoch using the Learning Rate Scheduler callbacks. However, as the learning curve reached a plateau in the last epoch, the learning curve decreased even more, and it reached 1.0e-8 for the last epoch.

The Adam algorithm, a simple-to-use, computationally efficient, and effective method, was used to optimize the categorical cross-entropy loss function to the minimum amount. The model was built and trained using the TensorFlow library v2.11 on python. The hyperparameters of the algorithms were tuned to achieve the desired outcome.

Results

Dataset description and statistical analysis:

This population contains information about 10015 participants with a mean age of 51.86±16.96 which was only reported for 9958 cases. The biological sex of the 5406 (54.1%) participants was Male, 4552 (45.5%) were Female, and for 57 (0.6%) of them, the gender was unknown. The skin lesion images were taken from various parts of the body. The back with 2192(21.9%) images, the lower extremities with 2077(20.7 %) images, the trunk with 1404(14.0%) images, the upper extremities with 1118(11.2%) images, and the abdomen with 1022(10.2) images were the 5 most involved parts of the body; other ports prevalence is described in Fig 4.
The diagnosis for each lesion was confirmed via a specific route; 5340 (53.3%) lesions were confirmed through histopathologic examinations, 3704(37.0%) lesions were confirmed by follow-up examinations, 902(9.0%) lesions were confirmed by expert consensus, and the ground truth for 69(0.7%) of lesions was confirmation by in-vivo confocal microscopy Fig 4.

This dataset includes a representative collection of all essential diagnostic categories in pigmented lesions, but these categories were highly imbalanced, as illustrated in Fig 5. Most lesions were melanocytic nevi with 6705(66.9%) images. The rest of the lesions were melanoma (N = 1113, 11.1%), benign keratosis (N = 1099, 11.1%), basal cell carcinomas (N = 514, 5.1%), actinic keratosis (N = 327, 3.3%), vascular skin lesions (142, 1.4%), and dermatofibroma (115, 1.1%).

To establish whether there is a meaningful relationship between age and the type of skin cancer, we conducted an ANOVA test to compare mean age between different classes of skin cancer. A one-way ANOVA demonstrated that there is a statistically significant main effect of skin cancer type on age (F [6, 9951] = 470.117, p < .001); see Fig 6. The mean age and standard deviation for each skin cancer type are described in Table 1. Post hoc analyses using the Scheffé post hoc criterion for significance indicated that the average age was significantly lower in the melanocytic nevi class than in all the other 6 skin cancer classes with p values less than 0.001. the results of the Post hoc analyses are available in supplementary materials.
Also, to compare age between different classes of sex, a one-way ANOVA was conducted, which demonstrated that the 5400 male participants with an average age of 54.54±17.15, the 4548 female participants with an average age of 48.71±16.15; and the 10 participants with unknown sex had an average age of 37.50±22.26. The effect of gender, therefore, was significant (F [2,9955] = 8.76, p < 0.001) in the age of the people with skin cancer in this dataset (see Fig 6).

ML model performance validation:

The Machine Learning model was trained and tested on a Google Colaboratory environment with an Intel(R) Xeon(R) 2.30GHz CPU processor and 13GB of RAM and NVIDIA Tesla T4 CUDA enabled GPU processor with CUDA 11.2 which has designed for high-performance computing, deep learning training and inference, machine learning, and data analytics. The model was created with Python 3.8.6, TensorFlow 2.11, Scikit-Learn 1.0.2, and Numpy as dependencies.
This section presents the experimental results of our model trained on the HAM10000 dataset. The model was trained for 19 epochs with a batch size of 32, and in every epoch, training accuracy, training loss, and validation accuracy, validation loss was calculated. We used an Adaptive Momentum (Adam) optimizer on Categorical Cross Entropy loss function with a dynamic learning rate (LR) starting from 0.001. For fine-tuning in order to make the optimizer converge faster and get closer to the global minimum of the loss function, the learning was set high in early epochs, and by getting closer to the global optimum, the learning rates decreased to take tiny steps toward the global optimum. Also, we used the ReduceLROnPlateau callback to reduce the LR even more if the validation loss did not improve after 3 epochs. The metrics and LR for each epoch are described in Table 2.
The model’s precision and Recall and F1 Scores are described and compared in Fig 7 based on each class. As demonstrated in this Figure, the model performs best in detecting melanocytic nevi lesions with an F1 score of 0.93. This performance difference between the classes is mainly due to the highly imbalanced classes of the dataset. As the model gets trained with lots of melanocytic nevi images (about 5364 melanocytic nevi images compared to 92 dermatofibroma images), inevitably, the model learns more patterns to detect this specific class and higher performance in this class. As demonstrated in Table 2, the final model’s accuracy on unseen images of the test set was 84.30%, with a loss of 0.4387.

The model’s worst predictions

After making the predictions for all test images, we sorted the wrong predictions by their inferred probability to find those images that the model guessed wrong with the highest confidence. This procedure helps find both the dataset and the model’s problems. It is possible that an image is incorrectly labelled and the model is actually doing right; for instance, in the top 20 most wrong predictions (Fig 8), we found 2 identical images, and one of them should be deleted from the dataset.

Web application:

We also developed a web application to enable researchers to assess our model by uploading their malignant skin lesion images to the application and getting the results instantly.
Please note that this application has no clinical or diagnostic value and is for research purposes only. Some randomized clinical trials should be conducted to find clinical evidence supporting the accuracy of these models. Our application is available online at https://tajerian.info/ham10000. This application is also available offline and installable on computer devices on both windows and Linux, which can be downloaded from GitHub (https://github.com/tajerian/ham10000-app).

Despite all limitations discussed before, this study represents an important step forward in the development of automated tools for skin cancer diagnosis. Future studies could build on your work by expanding the dataset to include more diverse populations and additional diagnostic categories. Additionally, it would be interesting to explore the use of other imaging modalities in combination with ML models to further improve diagnostic accuracy.
The trained model can serve as a valuable tool to assist dermatologists in their clinical practice, potentially improving the accuracy and speed of diagnosis. Future work may involve integrating the model into a user-friendly application or system to make it more accessible to healthcare providers and patients.

Abstract

Background:

Asthma is a chronic condition characterized by episodic wheezing, cough, and shortness of breath resulting from airway hyperresponsiveness and inflammation. Over 300 million people are affected worldwide, and its prevalence is increasing by 50% every decade. Assessing the quality of life in children with asthma is fundamental, as consistently poor health-related quality of life is associated with poorly controlled asthma. This study is aimed to evaluate and compare factors associated with HRQOL between healthy and children with asthma.

Methods:

In the current case-control study, 50 children aged 8–12 years with asthma (cases) enrolled at outpatient hospital clinics by a trained pediatric allergist/immunologist (A.P.) and matched with 50 healthy controls by age and sex. All enrolled subjects were interviewed utilizing the PedsQL questionnaire to assess health-related quality of life; also, patient demographics, including age, sex, and family income status, were obtained from a questionnaire.

Results:

100 children comprising 62 males and 38 females with a mean age of 9.63 ± 1.38 years, participated in this study. The average score of children with asthma was 81.63±9.38, and the average score for healthy participants was 89.58±7.91. We found that asthma was associated with a significant drop in health-related quality of life in this sample.

Conclusion:

the results indicated that the PedsQL score and its subscales, except social functioning, were significantly higher in children with asthma compared to healthy ones. Also, SABA use, nocturnal symptoms, and asthma severity are negatively related to the health-related quality of life.

Keywords:

Asthma; children; quality of life.

Methodes

Results

In this study, children participated in two groups: children with asthma and healthy controls. Each group had 50 participants comprising 31 males and 19 females. The demographic characteristics of the participants are presented in Table 1. The majority of participants were male (62%). The mean age of the participants was 9.63 ± 1.38 years. As shown in Table 1, the asthma and healthy control groups are homogeneous regarding age and sex. There was no significant difference in age (p = 0.94) or sex (p = 1) between children with asthma and healthy controls.
The pediatric allergist/immunologist determined the severity of asthma based on the EPR3 guidelines. Table 2 describes the distribution of asthma severity and its impairment and risk components. 36 patients (72%) had intermittent asthma severity, 11 (22%) were mild persistent, 3 (6%) were moderate persistent, and no one had severe persistent asthma.
As The clinical characteristics of the children with asthma is shown in Table 3, most of the participants used SABA More than 2 days per week (n=37, 74%), had zero or one Asthma exacerbations per year (n=47, 94%), had less than or equal to 2 Nocturnal symptoms per month (n=39, 78%), and used no corticosteroid (n=36, 72%).
The average PedsQL total Scale Score for all 100 participants was 85.63±9.5, The average score of children with asthma was 81.63±9.38, and the average score for healthy participants was 89.58±7.91. Figure 1 compares the PedsQL total Scale Score between asthma patients and healthy participants. The average PedsQL scores in healthy and children with asthma were compared using the independent sample T-test; as presented in Table 4, asthma significantly reduces the quality of life in children (P=0.001).

In order to identify chronic conditions, the Clinical Risk Groups (CRGs) classification was performed using the cutoff scores discussed in the methods section. We found that in children with asthma, 5 (10%) participants were classified as Major Chronic Conditions class versus 1 (2%) participant in healthy children, and 11 (22%) participants were classified as moderate chronic conditions in children with asthma versus 5 (10%) participants in healthy children, and 34 (68%) participants were classified as no or minor chronic conditions in children with asthma versus 44 (88%) participants in healthy children. The summary results of health condition classification by PedsQL score are shown in Table 5.
Four Generic Core subscales of PedsQL were also compared in healthy and children with asthma, and the results are presented in Table 4. As illustrated in Figure 2, all Generic Core subscales, except social functioning, are significantly lower in children with asthma.

The mean PedsQL score in children with asthma for SABA use was as follows: 91.00 for the “less than or equal to 2 days per week” group (N=1), 84.68±7.8 for the “More than 2 days per week” group (N=37), 72.40±6.6 for “Once every day” group (N=10), and 68.00±4.24 for “Many times every day” group (N=2). The relationship between the SABA use and the quality of life in children with asthma was investigated using the ANOVA test, and with P less than 0.001, we found that the use of the spray is significantly related to the quality of life (Figure 3).

The mean PedsQL score in children with asthma for nocturnal symptoms were as follows: 84.31±7.8 for the “less than or equal to 2 times per month” group (N=39), 74.13±9.57 for the “3 or 4 times per month” group (N=8), and 67.67±3.1 for “more than 4 times per month” group (N=3). The relationship between the nocturnal symptoms and the quality of life in children with asthma was investigated using the ANOVA test. With P less than 0.001, we found that the nocturnal symptoms are significantly related to the quality of life (Figure 4).

The mean PedsQL score in children with asthma for asthma severity were as follows: 85.31±7.3 for the “intermittent” group (N=36), 73.64±8.1 for the “mild persistent” group (N=11), and 67.67±3.1 for “moderate persistent” group (N=3). The relationship between the asthma severity and the quality of life in children with asthma was investigated using the ANOVA test. With P less than 0.001, we found that asthma severity is significantly related to the quality of life (Figure 5).

Conclusion

Abstract

Purpose:

Central nervous system involvement by Brucella species is the most morbid form of brucellosis disease. Studies on neurobrucellosis are scarce and limited to case reports and series. Brucella is unable to infect or harm neurons without the assistance of monocytes. This raises the question of whether ceftriaxone-based regimens are effective.

Methods:

The primary aim of this study was to identify, evaluate, and summarize the findings of all relevant individual studies in the past 30 years to help better understand the disease. To achieve this, a broad systematic search was undertaken to identify all relevant records. Epidemiological and clinical features of the disease were assessed by the pooled analysis of descriptive studies. Through a meta-analysis, the treatment period duration was compared between the ceftriaxone-based and oral regimens using Standardized mean differences to measure effect size.

Results:

448 patients were included in the Meta-analyses from 5 studies. A moderate positive effect was found for ceftriaxone-based regimens over oral treatments, and there was a significant difference between these two groups (SMD 0.428, 95% CI -0.63 to -0.22, I 2 =37.64). Neurobrucellosis has a different clinical picture in pediatric patients. The disease is less chronic in children. Fever, nausea and vomiting, fatigue, and abdominal pain were significantly more prevalent symptoms in children, and Convulsions, ascites, sensorineural hearing loss, and papilledema were significantly more prevalent signs in children than adults.

Conclusion:

It is recommended to initiate the treatment of neurobrucellosis with at least 14 days of IV ceftriaxone therapy in combination with oral therapy.

Keywords:

Neurobrucellosis, Meningitis, Ceftriaxone, Treatment, Children

Methodes

Results

Conclusion