Deep Adaptation Network for HILN Determination in Serum and Plasma
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automated diagnostic analysis systems face challenges in accurately determining the presence and degree of hemolysis, icterus, and lipemia in serum or plasma portions of specimens due to subjective manual inspection and occlusion by barcode labels, leading to potential human error and inefficiency.
Innovation Solution
A method using a deep adaptation network (DAN) with a classification network, transformation network, and segmentation/classification/regression network to process pixel data from images of specimen containers, distinguishing between serum and plasma portions and determining HILN categories without requiring extensive training on serum sample images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual visual inspection is used to determine HILN categories, then human judgment and interpretation can be applied, but subjectivity and human error increase, reducing reliability
Solution Approach 1:
The patent replaces the mechanical/visual inspection system with an automated machine vision system using image capture devices and deep learning networks. The classification network, transformation network, and segmentation/classification/regression network automatically analyze specimen images to determine HILN categories, eliminating subjective human judgment while maintaining or improving accuracy through consistent automated processing
Solution Approach 2:
The system enables self-service by allowing the machine vision system to autonomously perform HILN determination without requiring skilled laboratory technicians. The deep learning networks automatically process images, transform pixel data, and classify specimens, making the inspection process independent of human expertise while improving reliability through automated consistency
2Measurement precision
If extensive training data with annotated serum sample images is collected to improve classification accuracy, then measurement precision improves, but time and resource consumption increase significantly
Solution Approach 1:
The patent applies preliminary action by first collecting only plasma sample images for training the deep learning networks, rather than collecting both plasma and serum images. The transformation network then adapts the learned features to serum samples through image transformation, achieving accurate serum classification without the time-consuming process of annotating extensive serum training data
Solution Approach 2:
The system changes the parameter approach by transforming pixel data characteristics rather than collecting diverse training data for each sample type. The transformation network modifies plasma image features to match serum appearance characteristics, allowing the model to generalize from plasma training data to serum classification tasks, significantly reducing annotation requirements while maintaining precision
3Adaptability or versatility
If a single classification model is trained on both plasma and serum images, then versatility improves, but the complexity of training data collection and model development increases
Solution Approach 1:
The patent segments the classification task into distinct components: a classification network for identifying sample type, a transformation network for adapting features, and a segmentation/classification/regression network for determining HILN categories. This segmentation allows training on plasma samples only while achieving versatility for both plasma and serum classification through the transformation network's adaptation capability
Solution Approach 2:
The transformation network serves as an intermediary between the plasma-trained classification model and serum sample analysis. It transforms plasma image features into serum-appropriate features, enabling the model to handle both sample types without requiring separate training datasets, thus reducing training complexity while maintaining versatility
4Loss of information
If barcode labels are present on specimen containers for identification, then specimen tracking is improved, but occlusion of the specimen view increases, reducing measurement precision
Solution Approach 1:
The patent addresses occlusion by capturing images from multiple viewpoints and angles, adding dimensional diversity to the data collection. The system processes images taken from different orientations, allowing the deep learning networks to identify HILN categories even when barcode labels partially occlude the specimen from certain angles, maintaining measurement precision while preserving specimen identification
Data Source
Figure 1~2
Figure 3A~3B
Figure 3C~3D
AI summary
A method of characterizing a serum or plasma portion of a specimen in a specimen container provides an HILN (hemolysis, icterus, lipemia, normal) determination. Pixel data of an input image of the specimen container is processed by a classification network to identify whether the specimen contains plasma or serum. Pixel data representing a plasma sample are forwarded to a segmentation/classification/ regression network trained with plasma samples for HILN determination. Pixel data representing a serum sample are forwarded to a transformation network, wherein the serum sample pixel data is transformed into pixel data that matches pixel data of a corresponding previously-collected plasma sample by changing sample color, contrast, intensity, and/or brightness. The transformed serum sample pixel data are forwarded to the segmentation/classification/regression network for HILN determination. Quality check modules and testing apparatus configured to carry out the method are also described, as are other aspects.