TechBio News: Automated Image Diagnosis Europe (part II)
MetaphysicalCells: A newsletter about Science, Technology and AI Drug Discovery
Hi everyone 👋 and welcome back to another edition of MetaphysicalCells 🛸 only for paid subscribers (and thanks for all your support🫶).
This newsletter is all about Medical Imaging Data and Automated Image Diagnosis in Europe (part II). You can find the first part of this newsletter here: TechBio News: Automated Image Diagnosis Europe (part I).
Automated Image Diagnosis Europe (part II)
Cancer detection is a critical area of medical research and practice, as early diagnosis can significantly improve patient outcomes. Artificial intelligence (AI) techniques, particularly machine learning (ML) and deep learning (DL), have shown tremendous potential in automating cancer detection processes. These methods leverage vast amounts of medical data to identify patterns that may be difficult for human practitioners to detect manually (Radiology and Imaging, Pathology, Genomics).
To give an example, ML algorithms are widely used to analyze structured data such as patient demographics, lab results, and imaging features. Common ML techniques include: Support Vector Machines (SVM) used for classification tasks, such as distinguishing between benign and malignant tumors; Random Forests effective for handling high-dimensional data and identifying important features; and Logistic Regression a simple yet powerful method for binary classification problems like cancer detection. Moreover DL Models, especially convolutional neural networks (CNNs), excel at analyzing unstructured data like medical images for: Image Segmentation in order to identify regions of interest in medical scans (e.g., tumors in MRI or CT scans); Classification in order to decide whether an image contains signs of cancer; and Object Detection for locating specific abnormalities within an image. Finally, Natural Language Processing (NLP) is used to extract valuable information from unstructured text, such as clinical notes, pathology reports, and research papers, for identifying risk factors, predicting cancer progression, and summarizing patient histories.
AI-based cancer detection techniques (specifically focusing on DL models) like
🔹 DenseNet121 (designed for image classification and other tasks like segmentation. It belongs to the family of CNNs and is characterized by its unique connectivity pattern),
🔹 DenseNet201 (a CNN that is 201 layers deep),
🔹 Xception (a CNN that is 71 layers deep),
🔹 InceptionV3 (it is part of the Inception family of models, which are known for their ability to efficiently analyze complex visual data and provide accurate results. Inception is a family of CNN for computer vision, introduced by researchers at Google in 2014 as GoogLeNet, later renamed Inception v1),
🔹 MobileNetV2 (a CNN that is 71 layers deep),
🔹 NASNetLarge (is a CNN that is trained on more than a million images from the ImageNet database),
🔹 NASNetMobile (is a CNN that is trained on more than a million images from the ImageNet database),
🔹 InceptionResNetV2 (this network is 164 layers deep),
🔹 VGG19 (a CNN that is 19 layers deep), and
🔹 ResNet152V2 (a specific instance of ResNet, of 152 convolutional layers, a Global average pooling layer, a dense layer with 64 neurons, and a dense layer with SoftMax followed by a convolution layer),
were evaluated on image datasets for 7️⃣ seven types of cancer: brain, oral, breast, kidney, Acute Lymphocytic Leukemia, lung and colon, and cervical cancer (Automating cancer diagnosis using advanced deep learning techniques for multi-cancer image classification).
After all the models were rigorously evaluated, DenseNet121 achieved the highest validation accuracy as 99.94%, 0.0017 as loss, and the lowest Root Mean Square Error (RMSE) values as 0.036056 for training and 0.045826 for validation. These results revealed the capability of AI-based techniques in improving cancer detection accuracy, with 🔹DenseNet121🔹 emerging as the most effective model in this study.
DenseNet121 is an advanced CNN architecture specifically designed for image classification tasks, including Multi-Cancer detection, that consists of a series of densely connected blocks, where each layer within a block receives input from all preceding layers, promoting feature reuse and enhancing the flow of information.
In particular, DenseNet121 comprises 121 layers, allowing it to capture more complex features in medical images. Each dense block contains several convolutional layers followed by batch normalization and ReLU activation functions. The architecture uses a growth rate parameter, which determines how many feature maps are added at each layer, facilitating the efficient expansion of the model’s representational capacity. Transition layers are strategically placed between dense blocks to reduce the dimensionality of the feature maps through average pooling, thus controlling the model’s computational efficiency and helping to mitigate overfitting.
This architecture is particularly effective 🦸♂️ in Multi-Cancer detection as it enables the models to learn intricate patterns associated with different cancer types, resulting in improved classification accuracy and enhanced diagnostic capabilities in medical imaging applications.
Regarding cancer detection, a lung cancer detection imaging company in Europe is Contextflow GmbH.
▶️ Contextflow GmbH
Contextflow (Vienna, Austria) is a spin-off of the Medical University of Vienna (MUW), Technical University of Vienna (TU) and European research project KHRESMOI. Founded in July 2016, Contextflow is a comprehensive solution for all chest CT’s: lung 🫁 cancer, (Interstitial lung disease) ILD and (Chronic Obstructive Pulmonary Disease) COPD. Their computer-aided detection system 🔺ADVANCE Chest CT🔻 offers lung nodule detection and tracking of changes over time, quantitative and qualitative analysis of 8 key image patterns, and 3D image search for 19 image patterns.
By integrating new technology from partner RevealDx, nodules can be further characterized, detecting lung cancer earlier and reducing costs. Contextflow ADVANCE Chest CT includes a TIMELINE feature that even quantifies and visualizes nodules over time, allowing physicians to follow treatment progress easily and efficiently. Fully integrated into the radiologist's workflow, preliminary research showed that average reading time was 31% shorter when Contextflow was available for use (European Radiology July 2022).
On November 25, 2024, Contextflow announce an agreement with IKK Südwest, a forward-thinking health insurance company in Germany, that was facilitated through the Healthy Hub initiative and marks a pivotal moment in the integration of AI technology within healthcare systems, with the goal to improve patient outcomes by enabling efficient and early detection of lung cancer (contextflow Set to Secure Europe’s First Radiology AI Reimbursement Contract).
Specific to this new partnership, malignancy scoring from partner RevealDx (integrated into Contextflow ADVANCE Chest CT) plays an important role. In fact, this feature enables radiologists to detect lung cancer up to one year sooner with significantly higher accuracy.