The scan finishes at 4:17 in the morning, though no one is awake to see it happen. In a beige machine in a hospital basement in Ohio, a CT scanner completes its final rotation around an unconscious man’s skull. Two hundred individual slices of his brain solidify into data: a language no human can read, but that modern medical imaging has, almost everywhere, agreed to speak.
The man’s name is Robert Aldridge, an illustrative patient this book will follow rather than a documented case. He came in by ambulance a little under an hour ago after collapsing at a gas station. Somewhere upstairs, a resident is refreshing a screen, waiting.
...
That pipe has a name. DICOM.
The Prologue continues with how an image travels from the scanner to the radiologist, and how it can fail.
Open a DICOM file and the first thing you notice is that it does not look like a picture at all.
It looks like a ledger.
...
Inside the file, each of those facts about Robert Aldridge lives in a data element, and an element has more to it than a value. It begins with a tag: a pair of numbers, written like (0010,0010), that says which attribute this is. Tag (0010,0010) is the patient’s name, in every conformant implementation that carries that attribute. Tag (0008,0060) is the modality, and its values are short codes rather than words: CT, MR, US, CR. Alongside the tag, the element carries a short code for what kind of value follows (a person’s name, a date, a decimal number, a block of raw bytes) and a count of how many bytes the value occupies. The kind-code is called a value representation, and depending on which encoding a system is using it is either written into the data explicitly or left implicit and looked up in a table both sides are assumed to hold. Then, finally, the value.
...
Those objects are also organized, and the organization matters more than it sounds. A patient has studies. A study is one imaging exam: one visit, one clinical question, though a combined examination like a PET/CT will put series from more than one device inside it. Inside a study are series, each a group of images acquired together under the same settings, the scout view, the pre-contrast series, the post-contrast series. Inside a series are instances, and an instance is one stored object, usually a single image but not always, since later chapters will show reports, segmentations and measurements stored the same way. The standard’s name for each stored object is a SOP Instance, and the term turns up everywhere in this field. An instance is one actual object: this slice, this report, this segmentation. What it carries is a declaration of what kind of thing it is, written into a tag called the SOP Class UID, which points at the class the object belongs to. A receiving system reads that declaration first. Whether it will accept the object is settled separately, when the two systems open a connection and agree, class by class, on what they are willing to exchange, which is why a study can be refused over a connection that opened perfectly well. Study, series and instance each carry their own globally unique identifier. The patient level does not: a patient ID is issued by whoever assigned it and means nothing outside that authority’s domain, which is the gap the master patient index exists to close and a reason identity recurs as a problem through the rest of this book. Below the patient, uniqueness is guaranteed. That is how an archive holding ten million objects, asked for one study, knows it has the right one. It is also why identifiers that drift, get regenerated, or get reused cause the specific and maddening class of failure Chapter 8 describes.
The genius of all this is that it doesn’t matter which company built the machine that generated the file. A GE tag (0010,0010) and a Siemens tag (0010,0010) are identical. This is the entire problem from Chapter 1, solved, not by forcing every manufacturer to build identical machines, but by forcing every manufacturer to agree on a shared vocabulary for describing what their very different machines produce.
...
A radiologist examining a CT scan on her workstation is rarely looking at the entire acquired range of gray values at once. A standard display can only show a limited range of distinguishable brightness levels at any one time, far short of the thousands of stored values a 12- or 16-bit acquisition makes available. Instead, she uses software controls, commonly called “windowing,” to select which narrow slice of that full stored range gets mapped onto the screen’s limited display range, effectively zooming in, tonally, on the specific tissue she’s trying to evaluate. This is a display operation, not a change to the underlying data: not a single number in the file is altered. A radiologist examining lung tissue will select a window tuned to the subtle gray-value differences within air-filled lung. That same image, viewed moments later with a window tuned for dense bone, would render the lungs as a nearly featureless gray blur, all the clinically relevant detail compressed into a range too narrow for the screen to display distinctly.
Chapter 2 goes on to cover transfer syntaxes and compression, the size of modern studies, and the network operations (C-STORE, C-FIND, C-MOVE) that move images between systems. Figure 2, which shows the structure of a DICOM object, accompanies it in the book.
From DICOM: The Invisible Pipeline by Mark W. Carr. Copyright © 2026 Mark W. Carr. All rights reserved. These excerpts are provided for reading and may not be reproduced beyond brief quotation with attribution.
Robert Aldridge is an illustrative patient, not a documented case. See the Chapters page for what the rest of the book covers, and the home page for how to cite or obtain the book.