A Weakly Supervised Deep Generative Model for Complex Image Restoration and Style Transformation

Abstract

<p>The datasets for transforming autofluorescence images to the histochemically stained images were acquired from a human breast biopsy tissue and a human liver cancer tissue. Tissues of breast cancer and liver cancer were extracted surgically or through tissue biopsy. The tissues were formalin-fixed and paraffin-embedded (FFPE). Thin tissue slices, with a thickness of 4 µm, were sectioned and placed on a quartz slide. The tissue slices were deparaffined prior to imaging. The autofluorescence images were acquired from a wide-field inverted microscope equipped with a 10X/0.3 numerical aperture (NA) objective lens (Plan Fluorite, Olympus Corp.), an infinity-corrected tube lens (TTL-180-A, Thorlabs Inc.), and a monochrome scientific complementary metal-oxide-semiconductor camera (pco.panda 4.2, PCO. Inc.). A deep ultraviolet light-emitting diode of 265 nm (M265L4, Thorlabs Inc.) was used as an excitation light source because of its high absorption in cell nuclei [28], consequently providing high nuclear contrast without labels [29]. After acquiring the autofluorescence image, the same slide was stained with H&E and its bright-field images were captured using a whole-slide scanner equipped with a 20X/0.75 NA objective lens (NanoZoomer-SQ, Hamamatsu Photonics K.K). All human experiments were carried out in conformity with a clinical research ethics review approved by the Institutional Review Board of the Chinese University of Hong Kong/ New Territories East Cluster (reference number: 2021.597).</p>

Label-free and non-destructive pathology of human lung adenocarcinomas with ultraviolet single-plane illumination microscopy

“…However, MUSI significantly strengthens the contrast between nucleic acid and extracellular matrix with deep-UV absorption, thus showing higher consistency with the histochemical staining in routine pathological practice. Due to this reseason, our images can be easily virtually stained to mimic the appearance of the H&E-stained images via deep learning-based style transfer frameworks [25], [35] (Fig. S7).…”

Artificial intelligence in breast cancer histopathology

“…Techniques have been developed to bypass the tissue sectioning 70 and/or staining step 71 . The researchers used generative adversarial network (GAN) to recreate the H&E images from image captured from nonsectioned/stained tissue 72 . These images were assessed by pathologists in a blind evaluation test and were deemed comparable to their H&E counterparts 71 …”

“…71 The researchers used generative adversarial network (GAN) to recreate the H&E images from image captured from nonsectioned/stained tissue. 72 These images were assessed by pathologists in a blind evaluation test and were deemed comparable to their H&E counterparts. 71 The majority of the images are in standard RGB colour space, as most pathology systems were initially designed for visual analysis.…”