Deep Learning In Healthcare

Deep Learning In Healthcare

Presentation For the University of South Florida / Guidewell / Florida Blue Conference

Science

Early Advances:

  • Over the past few decades, mathematicians and computer scientists have ushered in a new era of scientific inquiry through the advent of machine learning
  • While early machine learning primarily targeted decision making processes and were quite limited in application, the scope of machine learning models has exploded due to the abundance of big datasets, both structured and unstructured. 
    • Healthcare has particularly proved to be a beneficiary of the data revolution, giving rise to the emerging interdisciplinary field of computational medicine.  
  • While medicine harbors immense loads of data, it is often held in dispersed forms, from images to video to audio. 
  • Complex data forms (such as images, videos, and audio clips) are known as “high dimensional” due to the copious amounts of data stored.
    • Performing traditional machine learning techniques is often computationally expensive
  • Deep learning, however, is well suited to efficiently tackle the computational analysis of these complex data forms seen in healthcare.
    • Convolutional neural networks have particularly demonstrated success in handling complex medical data (especially images). 
    • Alongside a series of convolutional and fully connected layers, filters such as pooling and activation functions transform the output accordingly. 
  • Researchers have tried their hand at applying convolutional neural networks to identify cancers
    • Khosravi et al. developed a convolutional neural network pipeline model to classify cancers in breast, lungs, and bladders. Their model proved quite successful, accurately detecting various types of cancer.
      • P. Khosravi, E. Kazemi, M. Imielinski, et al. Deep convolutional neural networks enable discrimination of heterogeneous digital pathology images. EBioMedicine, 27 (2018), pp. 317-328, 10.1016/j.ebiom.2017.12.026
    • Back in 2016, Litjens et al. similarly developed a convolutional neural network architecture in order to identify prostate cancer in specimens. 
      • G. Litjens, C.I. Sánchez, N. Timofeeva, et al.Deep learning as a tool for increased accuracy and efficiency of histopathological diagnosis. Sci. Rep., 6 (2016), Article 26286, 10.1038/srep26286
  • Witnessing the promise of deep learning for the early detection of cancers, deep learning fervor spread across the medical industry
  • As two of the most deadly cancers in America, colorectal cancer and breast cancer present untapped opportunities for mathematicians and medical professionals to join forces. 
    • Advances within the past year have proven the remarkable potency of deep learning models to identify these cancers at an early stage, and consequently offer a viable technological solution to preserve the health and well-being of men and women around the globe. 

Deep Learning in Colorectal Cancer

  • In 2020, Colorectal cancer accounted for 53,000 deaths in America, making it the second most fatal cancer behind lung cancer
  • Colonoscopy stands as the predominant test for diagnosis; however, colonoscopy is not always accurate and is considerably vulnerable to false negatives. 
    • Many cases of colorectal cancer appear months or years after a negative colonoscopy test; such cases are known as “postcolonoscopy of colorectal cancer”.
    • Needless to say, it is incredibly critical to develop a reliable means for colorectal cancer detection. 
  • Raman spectroscopy is a chemical analysis technique that analyzes the scattering of light in order to determine characteristics of tissue. 
    • Given the unique vibration patterns due to scattering and absorption, Raman spectroscopy can be employed to identify molecular species.
      • Intestinal tissue is concentrated in Raman active molecules, presenting a spectral measurement that is remarkably equivalent to the weighted spectral sum of all molecular species within the tissue
  • Comparing Raman spectra of cancerous tissues with normal tissues yields the specific Raman spectrum characteristics identifying cancerous tissues
  • The advent of machine learning has grown critical in such spectral analysis
    • Traditionally, Principal component analysis is used to reduce dimensionality and thus identify the primary characteristics
    • Additionally, k-nearest neighbors and support vector machines have been employed in attempts to classify spectral features
    • These aforementioned methods, unfortunately, fail to learn in-depth information because “they consider only the disperse input points while neglecting internal relations”
  • Deep Learning may prove effective in detection, due to the efficacy of convolutional neural networks in automatically capturing multidimensional information
  • Early work in medical deep learning pointed to its use in colorectal cancer detection; however, many studies lacked comprehensive datasets. 
    • An early seminal piece was Gala de Pablo et al.’s study of colorectal cell lines with Raman spectroscopy, using Principal Component Analysis to achieve 92.4% accuracy
      • J. Gala de Pablo, F. J. Armistead, S. A. Peyman, D. Bonthron, M. Lones, S. Smith, and S. D. Evans, “Biochemical fingerprint of colorectal cancer cell lines using label-free live single-cell Raman spectroscopy,” Journal of Raman Spectroscopy, vol. 49, no. 8, pp. 1323–1332, 2018
    • Due to the limited size of the dataset, the usage of PCA in this study — along with several others — hampered the success of the model.
  • Computer Scientists and Medical Professors at Zhejiang University in China developed a new method to detect colorectal cancer via Raman spectroscopy and deep learning.
    • The Raman spectra were collected of consenting patients with varying degrees of the cancer at the Second Affiliated Hospital of Zhejiang University School of Medicine.
    • The resulting dataset was divided into the conventional 80-10-10 training-validation-test split. 
    • The researchers conducted ample preprocessing and augmentation for baseline correction, signal denoising, and data normalization. 
    • Upon completion of data collection and preprocessing, the datasets were trained on a residual network (ResNet) model.
      • To prevent overfitting, ResNets employ shorter connections between input and outputs. 
      • There are two residual bottlenecks, each with batch normalizations, a ReLU activation layer, and convolutional layer sets.
      • The said convolutional layers each have a 1×3 kernel.
      • On top of all convolutional layers is a fully connected layer, upon which a sigmoid function was applied to compute the classification likelihood. 
    •  The researchers followed up their study by comparing the deep learning model to other machine learning models, including Support Vector Machines and Random Forest. 
      • The deep learning model achieved 98.5% accuracy, as opposed to 77.63% and 78.37% accuracy for SVM and Random Forest respectively. 
  • Ultimately, through careful analysis, Raman spectroscopy can indeed prove useful for colorectal cancer detection. 
  • Given that Raman spectroscopy produces high dimensional image dataset, deep learning models lend themselves well to automating the classification of the produced spectral images. 
  • This 2022 model demonstrates the efficacy of deep learning in colorectal cancer detection, and thus suggests that similar methods can be utilized to identify other cancers, such as breast cancer.  

Deep Learning in Breast Cancer: 

  • According to a 2018 study, breast cancer accounts for roughly 15% of cancer-linked deaths among women, thereby positioning the disease as one of the most fatal malignancies for females. 
    • As such, a reliable means for early-stage detection is imperative to save lives. 
  • Medical professionals diagnose breast cancer through several different techniques, including mammography and thermography.
    • While generally effective and inexpensive, mammography has historically been weak for patients with dense breast tissues, and may even lead to the onset of severe side-effects due to ionized radiation
    • Recently, thermography has risen to the forefront of medical technology, as it pertains to breast cancer screening. Largely due to its radiation-free and non-invasive nature, thermography can be efficiently used to detect early-stage breast cancer. 
  • Many early advances were made to optimize segmentation, develop meaningful feature extractions, and design accurate classifiers: 
    • As a preliminary step, it is critical to efficiently — and preferably automatically — segment the breast area away from the human body in thermography imagery
      • Mahmoudzdeh et al. employed hidden Markov models, BayesNet, and Random Forest to optimize breast segmentation techniques. This is indeed useful as a first step. 
        • Mahmoudzadeh E, Montazeri M, Zekri M and Sadri S. Extended hidden markov model for optimized segmentation of breast thermography images. Infrared Phys. Technol. 2015; 72: 19–28.
      • Other advances, including Ali et al. anld Gaber et al., involved alternative methods; however, these proposals suffered due to limited datasets.
        • Ali M, Sayed G, Gaber T, Hassanien A, Snasel V and Silva L. Detection of breast abnormalities of thermograms based on a new segmentation method.
        • Gaber T, Ismail G, Anter A, Soliman M., Ali N, Hassanien A, et al. Thermogram breast cancer prediction approach based on neutrosophic sets and fuzzy c-means algorithm.
    • Feature extraction seeks to extract particular features to obtain significant results. 
      • Araujo et al. reviewed datasets and performed statistical analysis to obtain a 85.7% sensitivity and 86.5% specificity. 
        • Arau´jo M, Lima R and Souza R. Interval symbolic feature extraction for thermography breast cancer detection. Expert Syst. Appl. 2014; 41: 6728–37.
      • De Santana et al. studied classification techniques and considered machine learning techniques such as artificial neural networks, decision trees, and Bayesian classifiers, obtaining a 76% accuracy
        • De Santana M, Pereira J, Da Silva F, De Lima N, De Sousa F, De Arruda G, et al. Breast cancer diagnosis based on mammary thermography and extreme learning machines. Res. Biomed. Eng. 2018; 34: 45–53.
      • Pramanik et al. took an approach of reducing dimensions with Principal Component Analysis and feeded the features into a Perceptron — notably a seminal machine learning technique developed at Cornell University —
        • They obtained 90.48% accuracy and 87.6% sensitivity. 
        • Pramanik S, Bhattacharjee D and Nasipuri M. Wavelet based thermogram analysis for breast cancer detection. In: Proceedings of the 2015 International Symposium on Advanced Computing and Communication, ISACC 2015; Silchar, India. 2016.
      • Deep Learning
        • Past publications  investigating the application of deep learning to breast cancer detection were promising but flawed, largely due to unbalanced datasets and inconsistencies in initial segmentation. 
        • Hence, we have the following….
  • Working in tandem with thermography, deep learning can amplify the efficacy of diagnosing breast cancer. 
    • Mathematicians at Suez Canal University and the British University of Egypt — Esraa A. Mohamed, Essam A. Rashed, Tarek Gaber, Omar Karam — developed a fully automatic breast cancer detection system in 2022. 
      • 1) U-Net Network — a convolutional neural network designed for biomedical image segmentation – is employed to automatically extract and isolate the breast from the remainder of the human body
      • 2) Apply a deep learning model, trained on an expansive dataset of breast tissue thermal images
    • U-Net Network for segmentation: 
      • Proposed by Olaf Ronnenberger at the University of Freiburg in Germany, the U-Net Network has historically performed well on smaller training datasets
        • Often, medical scientists are subjected to dealing with small training datasets due to the lack of comprehensive information available, hence rendering this network as a viable tool in medical technology.
      • The U-Net Network contains a contracting path and a expansive path
        • The former involves two 3×3 convolutions, passed through a ReLU which is then transformed with a 2×2 max-pooling operation. 
        • The latter passes the feature map through a 2×2 convolution and two 3×3 convolutions followed directly by a ReLU
      • Combining the contracting and expansive paths yields a U-shaped graph, hence the title “U-Net Network”. The graph can then be used to perform image segmentation
    • Deep Learning for classification: 
      • With nine layers — six convolutional layers and three fully connected layers —  a two-class CNN deep learning model was efficiently shown to classify breast tissue. 
        • The first layer filters the image with 64 7×7 kernels and a stride of 6 pixels
        • The output of the first layer — after max-pooling — is passed into the second layer that filters it with 128 kernels of size 3x3x64
        • The third layer held 256 kernels of size 3x3x128, without pooling
        • The fourth layer had 256 kernels of size 3x3x256, without pooling
        • The fifth layer consists of 256 kernels of size 3x3x256, without pooling
        • Max-pooling is applied to the output of the fifth layer, and then passed then passed through the sixth layer with 256 kernels of size 3x3x256
        • The model has two fully-connected layers of 1024 neurons
        • Finally, a ninth layer with the quantity of neurons equivalent to the number of classes; for binary classification, we have two classes. 
    • Results: 
      • The researchers employed Adaptive Moment Estimation with 30 epochs, a batch size of 60, and starting learning rate of 2.0*10-3 as an optimized algorithm for the computation.
        • Please note that additional settings and metrics were tested, slightly altering the results. 
      • 99.33% accuracy, 100% sensitivity and 98.67% specificity was achieved. 

Conclusions: 

  • Due to its ability to capture and process high dimensional datasets efficiently, deep learning has emerged as a leading technique for the early-stage detection of colorectal and breast cancers.
  • Nonetheless, the power of data also comes with a great responsibility — the responsibility of privacy.
    • Deep learning thrives off of massive, unstructured datasets; with its rise to prominence, big data will be shipped and stored by several stakeholders, potentially compromising sensitive data
  • Further advances in security are critical as a preventative measure to preserve the ethics of our new computational age. 
    • The introduction of blockchain into healthcare is an exciting opportunity to secure sensitive data. 
  • After all, machine learning began as a means to simulate human decision making— mathematics is deeply human at its core, and it is our duty as mathematicians, computer scientists, and medical professionals to engineer a future protecting our rights while advancing our fields.

Prepared by Alexander Fleiss & Avhan Misra

Deep Learning In Healthcare